何时可以对网络进行剪枝？多语言语音解析中中间神经元的研究

# When Can You Prune Your Network? A Study of Intermediate Neurons in Multilingual Speech Parsing
Source: [https://arxiv.org/html/2610.11520](https://arxiv.org/html/2610.11520)
Benjamin LecouteuxMaximin CoavouxAffiliation:Univ\. Grenoble Alpes, CNRS, Grenoble INP, LIG, 38000 Grenoble, FranceEmail:[first\.last@univ\-grenoble\-alpes\.fr](mailto:)

###### Abstract

End\-to\-end speech parsing, a task recently proposed, consists in predicting both the transcription and the syntactic tree for a spoken utterance\. Existing architectures for speech parsing often utilise intermediate neural networks\. In this work, we examine the effectiveness of intermediate neural networks \(NN\) for parsing, and, specifically, what role do they play\. We introduce a simpler end\-to\-end architecture for speech parsing, where we remove these intermediate NN units, reducing the parameters by 12%, while achieving comparable or better performance than prior method on both automatic speech recognition \(ASR\) and parsing\. We demonstrate that intermediate NN units help reduce therepresentational gapwhen the pre\-trained encoder is frozen\. We do a comprehensive evaluation of speech parsing on French, and medium\-low resource languages Slovenian and Naija\. We further investigate the impact of the training data size and intermediate layers of the pretrained speech encoder on speech parsing\.

## 1Introduction

End\-to\-end architectures for speech parsing\([Pupier et al\., 2022](https://arxiv.org/html/2610.11520#bib.bib18);[Kando et al\., 2024](https://arxiv.org/html/2610.11520#bib.bib14)\)connect the pre\-trained speech and the parsing models using intermediate neural network \(NN\) units like feedforward neural network \(FNN\), recurrent neural network \(RNN\), and long short\-term memory \(LSTM\) units\. These units increase the computational training time\. However, the impact of these units has not been yet studied for parsing\.

Even though dependency parsing \(Figure[1](https://arxiv.org/html/2610.11520#S1.F1)\) has nowadays fewer applications in NLP, the design of accurate syntactic parsers is still relevant for applications to data\-driven linguistics\([Hüll and Dobrovoljc, 2025](https://arxiv.org/html/2610.11520#bib.bib36);[Peck and Becker, 2024](https://arxiv.org/html/2610.11520#bib.bib37)\)whose analyses rely on syntactic annotations\. However, treebanks of spontaneous speech, arguably the most naturalistic source of linguistic data, are scarcer than written treebanks\.The parsing of spontaneous speech is currently not at a satisfactory level\. Hence, high\-quality speech parsers would open new perspectives for the study of syntax and for computational language documentation\. More generally, besides these motivations, our work subscribes to the current necessity for the NLP community to focus more on speech as the most naturalistic source of linguistic data\([Chrupała, 2023](https://arxiv.org/html/2610.11520#bib.bib30)\)\.

Wav2tree\([Pupier et al\., 2022](https://arxiv.org/html/2610.11520#bib.bib18)\)is an end\-to\-end architecture for spontaneous speech parsing where the pre\-trained speech model is fine\-tuned jointly with the downstream parsing model, with intermediary NN units inserted between the two models\. It takes speech as input, and produces both the transcription and the syntactic tree as the output\. Due to this joint fine\-tuning setup, the representations learned by the speech encoder, as well as those produced by the intermediate NN units, is optimised for the downstream parsing model\.

![Refer to caption](https://arxiv.org/html/2610.11520v1/figs/tree_IBA_04_Alaska-Pepe_MG__223.png)

Figure 1:Dependency parsing tree from the NaijaNSC treebank\.en: Ala\-… Alaska Pepper was shocked\. The two lines under the wordform line are the sequence of POS and lemma tags respectively\.Our contributions are the following:

1. 1\.We examine the role of intermediate neural network \(NN\) units in Wav2tree: - •we introduce a new end\-to\-end architecture,Simplified Wav2tree\(Simplified W2T\), for parsing speech to a dependency tree\. It simplifies Wav2tree by removing these intermediate NN units, while generally matching or improving performance with greater efficiency\.111[https://github\.com/minniekabra/SpeechParsing](https://github.com/minniekabra/SpeechParsing) - •When the pre\-trained model \(encoder\) is frozen within a joint parsing architecture, its representation remains optimised for the pre\-training objective, and not for the downstream parsing task\. We refer to this phenomenon as therepresentational gapbetween the encoder and the parsing model\. We demonstrate that intermediate neural network units help reduce thisrepresentational gapbetween the frozen encoder and the parsing model\.
2. 2\.We carry out a comprehensive evaluation of dependency parsing of speech for the spoken treebanks of Naija, Slovenian, and French: - •We compare Simplified Wav2tree with baseline Wav2tree\. - •We examine the performance of dependency parsing using representation from intermediate Wav2Vec2[Baevski et al\. \(2020\)](https://arxiv.org/html/2610.11520#bib.bib41)layers as the input to both ASR and parsing model in the Simplified architecture\. The performance of parsing remains stable or improves, increasing the parameter efficiency of the architecture\. - •We do a comparison with the pipeline approach and gold transcriptions\. Except for Slovenian, the end\-to\-end parsing performs better than the pipeline\.
3. 3\.We analyze the impact of scaling training data on dependency parsing of speech\. Through a two\-stage experimental setup, we assess the amount of speech data required by the parsing model to attain a near\-optimal performance with the Simplified architecture\.
4. 4\.We present language\-specific analyses for Naija and Slovenian\. For Naija, we assess the usefulness of segmentation annotations for parsing\. Our experiments show that they do not lead to improvements as they do when parsing transcriptions\. For Slovenian, we release the code[1](https://arxiv.org/html/2610.11520#footnote1)to produce filteredUD\_Slovenian\-SSTv2\.16 with a better speech\-transcription alignment\.

DatasetSectionDurationTreesSpeakersTrain6\.78h7,26867NaijaDev0\.87h99010Test0\.91h97211Train4\.88h3,474464SlovenianDev0\.67h40080Test0\.99h35836Train1\.34h1,265ParisStoriesDev0\.58h640N/ATest0\.59h631Train122\.1h158,778OrféoDev11\.57h15,0002000\+Test11\.62h15,000

Table 1:Data statistics for the Naija\-NSC treebank, Slovenian SST, ParisStories, and Orféo
## 2Related work

Historically, speech parsing has mostly been addressed through parsing transcriptions, be the gold transcriptions or predicted transcriptions\([Béchet et al\., 2014](https://arxiv.org/html/2610.11520#bib.bib16), e\.g\. \)\. Early work on speech parsing aimed at jointly detecting \(and sometimes removing\) disfluencies and parsing\([Charniak and Johnson, 2001](https://arxiv.org/html/2610.11520#bib.bib6);[Jørgensen, 2007](https://arxiv.org/html/2610.11520#bib.bib5);[Rasooli and Tetreault, 2013](https://arxiv.org/html/2610.11520#bib.bib8);[Rasooli and Tetreault, 2014](https://arxiv.org/html/2610.11520#bib.bib22);[Honnibal and Johnson, 2014](https://arxiv.org/html/2610.11520#bib.bib9);[Yoshikawa et al\., 2016](https://arxiv.org/html/2610.11520#bib.bib10);[Jamshid Lou et al\., 2019](https://arxiv.org/html/2610.11520#bib.bib21)\)\. The underlying motivation for removing the disfluencies is often to use the output tree as input to downstream NLP tools trained on written text and reduce the domain shift between speech and text\. Prior work also proposed to include audio features in text parsers, as an extra source of information to improve disambiguation\([Kahn et al\., 2005](https://arxiv.org/html/2610.11520#bib.bib15);[Dreyer and Shafran, 2007](https://arxiv.org/html/2610.11520#bib.bib17);[Pate and Goldwater, 2013](https://arxiv.org/html/2610.11520#bib.bib11);[Tran and Ostendorf, 2021](https://arxiv.org/html/2610.11520#bib.bib23)\)\. In contrast, instead of relying on transcriptions, we address speech parsing directly from the speech signal and perform automatic speech recognition jointly, in line with[Pupier et al\. \(2022\)](https://arxiv.org/html/2610.11520#bib.bib18)\.

Only a few publications have dealt with speech parsing without relying on transcriptions\([Pupier et al\., 2022](https://arxiv.org/html/2610.11520#bib.bib18);[Pupier et al\., 2024](https://arxiv.org/html/2610.11520#bib.bib19);[Kando et al\., 2024](https://arxiv.org/html/2610.11520#bib.bib14)\)\. They evaluate either on the French Orféo treebank\([Benzitoun et al\., 2016](https://arxiv.org/html/2610.11520#bib.bib4)\)for the first two, or Orféo and the English Switchboard treebank\([Godfrey et al\., 1992](https://arxiv.org/html/2610.11520#bib.bib7)\)for the last one\. This paper focuses instead on lower\-resource languages: Naija and Slovenian\. A second major novelty compared to these three publications is that we introduce a much simpler deep learning architecture\.

Our work is also connected to[Lai et al\. \(2023\)](https://arxiv.org/html/2610.11520#bib.bib12);[Tseng et al\. \(2023\)](https://arxiv.org/html/2610.11520#bib.bib13), who investigate grammar inference from the speech signal and visual cues on a multimodal dataset of images with read captions\. However, the data they use consists of captions read aloud by speakers\. Texts that are read aloud do not feature the phenomena that make speech recognition and analysis difficult \(disfluencies, interruptions, overlapping speakers\)\. Moreover, their syntax is similar to that of written texts\. As a result, there is a wide domain shift between spontaneous and read speech\. Our work focuses instead on corpora featuring realistic interactions\.

## 3Datasets

We conduct a multilingual evaluation of Simplified W2T on three spoken treebanks from Universal Dependencies\([Nivre et al\., 2020](https://arxiv.org/html/2610.11520#bib.bib24)\), covering three languages\. The multilingual evaluation here refers to training and evaluating the architecture separately on these three languages\. We additionally use the French Orféo treebank\([Benzitoun et al\., 2016](https://arxiv.org/html/2610.11520#bib.bib4)\)to assess the impact of scaling the training data on dependency parsing\. The statistics of the datasets is presented in Table[1](https://arxiv.org/html/2610.11520#S1.T1)\.

1. 1\.Naija\- NaijaSynCor \(Naija\-SNC\) treebank\([Caron et al\., 2019](https://arxiv.org/html/2610.11520#bib.bib1)\), released as part of Universal Dependencies\([Nivre et al\., 2020](https://arxiv.org/html/2610.11520#bib.bib24)\), was manually annotated using the Surface\-Syntactic Universal Dependency\([Gerdes et al\., 2018](https://arxiv.org/html/2610.11520#bib.bib40), SUD\)annotation scheme\. Naija, officially named Nigerian Pidgin, is an English\-based creole language, whose lexicon is influenced by Nigerian languages such as Yoruba, Hausa, and Igbo\. The treebank recordings span life stories, speeches, radio programs, free conversations, cooking recipes, comments on the current state of affairs, etc\.
2. 2\.Slovenian\- Slovenian SST\([Dobrovoljc and Nivre, 2016](https://arxiv.org/html/2610.11520#bib.bib31)\)was manually annotated using the Universal Dependencies \(UD\) annotation scheme\. Its recordings cover multiple categories, including public informative and educational content, public entertainment, non\-public non\-private interactions, and non\-public private conversations among friends or family members\.
3. 3\.
4. 4\.French Orféo treebank333Around 5% of the Orféo treebank has been manually annotated \(gold\), the rest of the corpus is automatically annotated \(good quality silver data\)\- Orféo is the only publicly available spoken treebank large enough for scaling experiments\. It has multiple sub\-corpora, and the recordings span work meetings, interviews, public meetings, storytelling, narratives, and telephone conversations\.

We observed speech noise in the Slovenian\-SST dataset which we filtered out, and enhanced speech\-transcription misalignment\.We provide further details in Appendix[C](https://arxiv.org/html/2610.11520#A3)\.

![Refer to caption](https://arxiv.org/html/2610.11520v1/figs/ExistingWav2tree.drawio.png)

\(\(a\)\)Baseline Wav2tree\([Pupier et al\., 2024](https://arxiv.org/html/2610.11520#bib.bib19)\)
![Refer to caption](https://arxiv.org/html/2610.11520v1/figs/SimplifiedWav2tree.drawio.png)

\(\(b\)\)Simplified Wav2tree \(ours\)

Figure 2:Comparison of the original architecture\([Pupier et al\., 2024](https://arxiv.org/html/2610.11520#bib.bib19)\)and our proposed architecture\.
## 4Background: Wav2tree Architecture

Wav2Tree\([Pupier et al\., 2022](https://arxiv.org/html/2610.11520#bib.bib18)\)is an end\-to\-end dependency parsing model whose only input is the raw speech signal\. Its architecture is illustrated in Figure[2\(a\)](https://arxiv.org/html/2610.11520#S3.F2.sf1), and performs speech recognition and parsing jointly\. The Wav2tree architecture is composed of three modules: \(i\) a speech encoder that computes a vector representation for each frame in the speech signal, \(ii\) an ASR module that outputs both a transcription and the frame boundaries of each predicted token, \(iii\) a parsing module that uses the word boundaries to compute audio word embeddings and run a standard parsing algorithm on the sequence of embeddings\. These modules are connected by three intermediate NN units\.

The audio signal𝐬\\mathbf\{s\}is first passed to a pre\-trained speech encoder \(Wav2Vec2, Eq\.[1](https://arxiv.org/html/2610.11520#S4.E1)\) to obtain a sequence of frame vectors𝐅\(1\)∈ℝℓ×f\\mathbf\{F\}^\{\(1\)\}\\in\\mathbb\{R\}^\{\\ell\\times f\}, whereℓ\\ellis the length of the signal in number of frames andffis the output size of Wav2Vec2 representations\. Then the obtained matrix is passed through the 1stintermediate NN unit \- Feedforward neural network \(feature encoder, Eq\.[2](https://arxiv.org/html/2610.11520#S4.E2)\)\.

𝐅\(1\)\\displaystyle\\mathbf\{F\}^\{\(1\)\}=Wav2vec⁡\(𝐬\),\\displaystyle=\\wavto\(\\mathbf\{s\}\),\(1\)𝐅\(2\)\\displaystyle\\mathbf\{F\}^\{\(2\)\}=FNN⁡\(𝐅\(1\)\)\.\\displaystyle=\\fnn\(\\mathbf\{F\}^\{\(1\)\}\)\.\(2\)The speech recognition module applies a Connectionist temporal classification layer\([Graves et al\., 2006](https://arxiv.org/html/2610.11520#bib.bib25), CTC\)to𝐅\(2\)\\mathbf\{F\}^\{\(2\)\}\. The granularity of the CTC prediction is at the character level: each frame is tagged with a character \(or a special blank character representing a separation between two words\)\.

To obtain transcriptions from CTC predictions, Wav2tree uses a greedy decoding algorithm444[\{https://github\.com/Pupiera/ParseBrain/blob/master/parsebrain/speechbrain\_custom/decoders/ctc\.py\}](https://{https//github.com/Pupiera/ParseBrain/blob/master/parsebrain/speechbrain_custom/decoders/ctc.py%7D)\(Eq\.[3](https://arxiv.org/html/2610.11520#S4.E3)\)\. It is a modification of the greedy decode from the SpeechBrain library\([Ravanelli et al\., 2021](https://arxiv.org/html/2610.11520#bib.bib29)\), which outputs a single transcription\. Along with the transcriptiont^\\hat\{t\}, the decoding algorithm outputs a mappinga^\\hat\{a\}of the frames of Wav2Vec2 to the words, based on the character tagged on each frame\. We provide an example of a mapping in Figure[5](https://arxiv.org/html/2610.11520#A1.F5)in Appendix\. Thepredicted word boundaries\(start and end frames for each transcribed word\) are extracted from this mapping\.

t^,a^\\displaystyle\\hat\{t\},\\hat\{a\}=CTC⁡\(𝐅\(2\)\)\.\\displaystyle=\\ctc\(\\mathbf\{F\}^\{\(2\)\}\)\.\(3\)
The 2ndintermediate NN unit LSTM then uses these predicted word boundaries to compute the ‘audio’ word embedding i\.e\., fixed\-sized vectors for each predicted wordwkw\_\{k\}\(k∈\{1,…​n\}k\\in\\\{1,\\dots n\\\}\) with boundary indices\(ik,jk\)\(i\_\{k\},j\_\{k\}\)\. This process is akin to using a character LSTM to compute textual word embedding based on their internal structure\.\([Ling et al\., 2015](https://arxiv.org/html/2610.11520#bib.bib26);[Plank et al\., 2016](https://arxiv.org/html/2610.11520#bib.bib27)\)\. Instead of a sequence of characters, spoken words are a sequence of frames \(the output of thefeature encodermodule\)\.

𝐰k\\displaystyle\\mathbf\{w\}\_\{k\}=LSTM\(𝐅⋅,ik:jk\(2\)\)\.\\displaystyle=\\lstm\(\\mathbf\{F\}^\{\(2\)\}\_\{\\cdot,i\_\{k\}:j\_\{k\}\}\)\.\(4\)Next, Wav2tree computes contextualised audio word embedding by passing the sequence of embeddings\(𝐰1,…,𝐰n\)\(\\mathbf\{w\}\_\{1\},\\dots,\\mathbf\{w\}\_\{n\}\)through the 3rdintermediate NN unit \- a bidirectional LSTM \(Eq\.[5](https://arxiv.org/html/2610.11520#S4.E5)\):

\(𝐡1,…,𝐡n\)\\displaystyle\\mathbf\{\(\}\\mathbf\{h\}\_\{1\},\\dots,\\mathbf\{h\}\_\{n\}\)=bi\-LSTM⁡\(𝐰1,…,𝐰n\)\.\\displaystyle=\\bilstm\(\\mathbf\{w\}\_\{1\},\\dots,\\mathbf\{w\}\_\{n\}\)\.\(5\)The resulting word\-level representations are passed to a generic dependency parsing algorithm, either a seq2label strategy\([Strzyz et al\., 2019](https://arxiv.org/html/2610.11520#bib.bib3)\)or a biaffine architecture\([Dozat and Manning, 2017](https://arxiv.org/html/2610.11520#bib.bib2)\)\. The whole Wav2tree model is trained end\-to\-end with a multitask objective \(ASR and parsing\) to reduce the error propagation between the two tasks\.

## 5Experiments

The first set of experiments we conduct aims to answer the following research questions:

- •RQ1: What is the role of the intermediate neural network units in end\-to\-end parsing, and when do they add value?
- •RQ2: How is speech parsing accuracy affected by the size of the training data across two successive setups \(i\) when the ASR model is fine\-tuned jointly with the parsing model, and \(ii\) when a pre\-fine\-tuned ASR model is used to control for the effect of ASR accuracy\.

The second set of experiments provides a comprehensive evaluation of speech parsing across the three spoken treebanks\.

### 5\.1RQ1: Role of Intermediate NN Units

When the pre\-trained model is fine\-tuned jointly with the parsing model, its representations become aligned with the downstream parsing task\. Building on this observation, we propose two hypotheses regarding the role of intermediate neural network units in the parsing architecture\.

- •H1: When the pre\-trained model is fine\-tuned jointly with the parsing model, the intermediate NN units do not add an incremental value\.
- •H2: However, when the pre\-trained model is frozen, the intermediate NN units help to reduce therepresentational gapbetween the frozen pre\-trained model and the parsing model\.

#### 5\.1\.1\(H1\) Simplified Architecture

In Wav2tree \(Figure[2\(a\)](https://arxiv.org/html/2610.11520#S3.F2.sf1)\), the speech encoder is fine\-tuned jointly with the CTC layer, the parsing model, and three intermediate NN units\. We simplify the Wav2tree architecture by removing these intermediate NN units\.

1. 1\.We removed thefeature encodermodule\. The representations from Wav2Vec2, learned via a self\-supervised transformer\-based architecture, largely encapsulate cross\-segment dependencies without requiring further transformation\. Therefore, in practice, we apply CTC directly to𝐅\(1\)\\mathbf\{F\}^\{\(1\)\}\.
2. 2\.For word\-level representation, we directly use the representation from the speech model, by averaging all frames within each word\. These representations are derived from a transformer\-based encoder with self\-attention, and thus integrate contextual information across the frames for a word\. This suggests that the LSTM unit \(Eq\.[4](https://arxiv.org/html/2610.11520#S4.E4)\) is redundant in this setting\.
3. 3\.We hypothesize that the bi\-LSTM module is used to align the speech encoder representation with the parsing model\. Since the architecture is fine\-tuned end\-to\-end, i\.e\., the weights of the speech encoder are fine\-tuned for the downstream parsing task, the bi\-LSTM \(Eq[5](https://arxiv.org/html/2610.11520#S4.E5)\) is not required in this setting\.

The last two simplifications consist of feeding the sequence of vectors\(𝐅mean​\(i1,j1\)\(1\),…​𝐅mean​\(in,jn\)\(1\)\)\(\\mathbf\{F\}^\{\(1\)\}\_\{\\text\{mean\}\(i\_\{1\},j\_\{1\}\)\},\\dots\\mathbf\{F\}^\{\(1\)\}\_\{\\text\{mean\}\(i\_\{n\},j\_\{n\}\)\}\)to the parser, effectively bypassing Eq\.[4](https://arxiv.org/html/2610.11520#S4.E4)and[5](https://arxiv.org/html/2610.11520#S4.E5)\.

The audio word embeddings\(𝐡1,…,𝐡n\)\(\\mathbf\{h\}\_\{1\},\\dots,\\mathbf\{h\}\_\{n\}\)in our architecture are computed by the three following equations \(that replace Eq\.[1](https://arxiv.org/html/2610.11520#S4.E1)to[5](https://arxiv.org/html/2610.11520#S4.E5)\):

𝐅\(1\)\\displaystyle\\mathbf\{F\}^\{\(1\)\}=Wav2vec⁡\(𝐬\),\\displaystyle=\\wavto\(\\mathbf\{s\}\),\(6\)t^,a^\\displaystyle\\hat\{t\},\\hat\{a\}=CTC⁡\(𝐅\(1\)\),\\displaystyle=\\ctc\(\\mathbf\{F\}^\{\(1\)\}\),\(7\)\(𝐡1,…,𝐡n\)\\displaystyle\\mathbf\{\(\}\\mathbf\{h\}\_\{1\},\\dots,\\mathbf\{h\}\_\{n\}\)=\(𝐅mean​\(i1,j1\)\(1\),…​𝐅mean​\(in,jn\)\(1\)\),\\displaystyle=\(\\mathbf\{F\}^\{\(1\)\}\_\{\\text\{mean\}\(i\_\{1\},j\_\{1\}\)\},\\dots\\mathbf\{F\}^\{\(1\)\}\_\{\\text\{mean\}\(i\_\{n\},j\_\{n\}\)\}\),\(8\)wherea^=\[\(i1,j1\),…​\(in,jn\)\]\\hat\{a\}=\[\(i\_\{1\},j\_\{1\}\),\\dots\(i\_\{n\},j\_\{n\}\)\]as previously mentioned, and𝐅mean​\(ik,jk\)\(1\)\\mathbf\{F\}^\{\(1\)\}\_\{\\text\{mean\}\(i\_\{k\},j\_\{k\}\)\}is the vector representation of the mean of all frames for a wordkk\.

In all, the output of speech representation from the speech encoder now directly serves as the input to the parsing model \(see Figure[2\(b\)](https://arxiv.org/html/2610.11520#S3.F2.sf2)\)\. With this change, the parameter count is reduced by12%in the simplified architecture, considering the hyperparameters used by[Pupier et al\. \(2024\)](https://arxiv.org/html/2610.11520#bib.bib19)\.

#### 5\.1\.2\(H2\) When Do Intermediate NN Units Add Value?

We evaluate hypothesis H2 using a text parsing model\. The text parsing architecture we use is the same as the Simplified architecture for speech parsing, except that the speech encoder is replaced by a pretrained text encoder\. To study and isolate the impact of intermediate NN units, we feed gold transcriptions into the text parser\. We do two experiments across all three spoken treebanks:

1. 1\.Joint fine\-tuning: We first evaluate the text parsing performance when the architecture is fine\-tuned end\-to\-end, with and without an intermediate NN unit \(LSTM\)\.
2. 2\.Frozen pre\-trained text model: Next, we freeze the pre\-trained text model \(Figure[3](https://arxiv.org/html/2610.11520#S5.F3)\), and evaluate the text parsing performance with and without an intermediate NN unit\. ![Refer to caption](https://arxiv.org/html/2610.11520v1/figs/frozen.drawio.png) \(\(a\)\)No intermediate NN unit\. ![Refer to caption](https://arxiv.org/html/2610.11520v1/figs/frozen_lstm.drawio.png) \(\(b\)\)With LSTM Unit Figure 3:Text parsing with frozen text encoder \- with and without intermediate NN unit\.

We note that it is not possible to conduct the same experiment for speech parsing, as the underlying speech model \(Wav2Vec2\) requires fine\-tuning for the target language\.

### 5\.2RQ2: Scaling Training Data

We use the two largest spoken treebanks \(Naija\-SNC, 7h; Orféo, 122h\) to assess the effect of the training set size\. The training data was partitioned into five subsets for this experiment: 20%, 40%, 60%, 80%, and 100%\. With the simplified speech parsing architecture, we assess the amount of speech required to achieve an optimal performance\. We conduct two\-stage experiments for it:

1. 1\.Baseline setup: We ran the Simplified W2T, without any modifications, across all data partitions\.
2. 2\.Frozen Speech encoder: Since the input to parsing is the learned representation produced by the ASR model \(speech encoder & CTC tagging\), ASR accuracy directly affects parsing performance\. To isolate and precisely evaluate the effect of scaling training data on parsing, we use a pre\-fine\-tuned ASR system\. The pre\-fine\-tuned model is obtained from step[1](https://arxiv.org/html/2610.11520#S5.I5.i1), when trained with 100% of the training data\. Extending the finding from Section[5\.1\.2](https://arxiv.org/html/2610.11520#S5.SS1.SSS2), we use an intermediate LSTM unit to align representation from the frozen ASR model to the parsing model\.

### 5\.3Dependency Parsing Evaluation

##### Middle Layers of Wav2Vec2

On the simplified architecture and for all three spoken treebanks, we assess the impact of the middle layers of the Wav2Vec2 on dependency parsing, i\.e\., using the representation from intermediate layers as the input to both ASR and parsing\. This is motivated by the observations of[Shen et al\. \(2023\)](https://arxiv.org/html/2610.11520#bib.bib28)and[Dugonjić et al\. \(2024\)](https://arxiv.org/html/2610.11520#bib.bib20)who found through probing experiments that syntax is best encoded in the middle layers of the pre\-trained speech models\.

##### Baseline experiments

On all three spoken treebanks, we checked the performance of text\-based dependency parsing using a PIPELINE and a GOLD approach:

- •PIPELINE: The input to the text model is the predicted transcriptions from the pre\-trained speech encoder of the Simplified architecture
- •GOLD: The inputs to the text parser are gold transcriptions\. This experiment is meant as an upper bound \(ideal case with perfect ASR\)\.

### 5\.4Segmentation Annotations in Naija

The Naija dataset contains syntactic and prosodic segmentation annotations materialized by punctuation marks that are included as tokens in the dependency trees \(see the tree in Figure[1](https://arxiv.org/html/2610.11520#S1.F1)for an illustration\)\.[Caron et al\. \(2019\)](https://arxiv.org/html/2610.11520#bib.bib1)found these markups useful for parsing, with gold transcriptions as the input\. We found this is not the case when parsing the audio signal \(details in Appendix[B](https://arxiv.org/html/2610.11520#A2)\)\.

ModelWav2Vec2Intermediate NNWER ↓UPOS ↑UAS ↑LAS ↑ParametersBaseline W2T24 layersyes31\.477\.870\.262\.7315Mw\+13Mi\+4\.5MpSimplified W2T24 layersno31\.879\.169\.161\.8315Mw\+5MpSimplified W2T21 layersno31\.578\.568\.561\.3278Mw\+5MpPIPELINE24 layersno31\.877\.968\.061\.2315Mw\+5Mp\+559MrGOLD\-no0\.096\.988\.083\.9559Mr\+5Mp

\(\(a\)\)Evaluation on the Naija\-NSC treebank\.
ModelWav2Vec2Intermediate NNWER ↓UPOS ↑UAS ↑LAS ↑ParametersBaseline W2T24 layersyes33\.273\.046\.637\.7315Mw\+13Mi\+4\.5MpSimplified W2T24 layersno33\.076\.845\.738\.9315Mw\+5MpSimplified W2T18 layersno34\.376\.747\.440\.7240Mw\+5MpPIPELINE24 layersno32\.980\.256\.350\.9315Mw\+5Mp\+559MrGOLD\-no0\.097\.874\.971\.3559Mr\+5Mp

\(\(b\)\)Evaluation on the Slovenian treebank\.
ModelSpeech\_large\_frIntermediate NNWER ↓UPOS ↑UAS ↑LAS ↑ParametersBaseline W2T16 layersyes38\.272\.158\.450\.9313Mg\+13Mi\+4\.5MpSimplified W2T16 layersno37\.673\.659\.953\.3313Mg\+5MpSimplified W2T11 layersno38\.173\.560\.754\.2250Mg\+5MpPIPELINE16 layersno37\.571\.755\.948\.7313Mg\+5Mp\+559MrGOLD\-no0\.096\.176\.471\.9559Mr\+5Mp

\(\(c\)\)Evaluation on the ParisStories treebank\.

Table 2:Evaluation ontestdataset with settings described in Section[5\.5](https://arxiv.org/html/2610.11520#S5.SS5)\. Parameter counts:wWav2vec,gSpeech\_Large\_fr\_114K,ffeedforward \+ LSTM,pparsing module,rRoberta module\. W2T refers to Wav2tree\. Best scores across end\-to\-end speech parsing highlighted in bold\. Underlined refers to statistical significant resultsText parsing architectureIntermediate NNLASLASLAS\# of parameters\(LSTM\)\(Naija\)\(Slovenian\)\(ParisStories\)\(fine\-tuned\)Joint fine\-tuningno83\.971\.372\.0559Mr\+5MpJoint fine\-tuning \+ LSTMyes81\.766\.169\.8559Mr\+5Mp\+16MiFrozen pre\-trained text modelno51\.540\.944\.85MrFrozen pre\-trained text model \+ LSTMyes74\.655\.462\.95Mp\+16Mi

Table 3:Comparative analysis for RQ1 \- LAS on the test dataset with and without intermediate NN unit for all 3 treebanks with experiments described in Section[5\.1\.2](https://arxiv.org/html/2610.11520#S5.SS1.SSS2)\. Best scores across sections highlighted inbold\.
### 5\.5Training and Evaluation Details

##### End\-to\-End speech parsing

We usewav2vec2\-xls\-r\-300m,555[https://huggingface\.co/facebook/wav2vec2\-xls\-r\-300m](https://huggingface.co/facebook/wav2vec2-xls-r-300m)a multilingual Wav2Vec2 pre\-trained acoustic model, as the speech encoder for both Naija\-NSC treebank and Slovenian SST\. It has Slovenian as medium\-resource language \(11k hours\), and Yoruba & Hausa \(languages associated with Naija\) as low\-resource language \(75 hours each\) in its pre\-training data\. For UD French ParisStories, we useSpeech\_Large\_fr\_114K, a data2vec\([Baevski et al\., 2022](https://arxiv.org/html/2610.11520#bib.bib34)\)pre\-trained speech model trained on 114k hours of French speech\([Le et al\., 2026](https://arxiv.org/html/2610.11520#bib.bib39)\)\.

The parsing module is a biaffine graph\-based parser\([Dozat and Manning, 2017](https://arxiv.org/html/2610.11520#bib.bib2)\)\. The baseline Wav2tree results are produced with the original Wav2tree implementation\.

##### Text parsing

We usexlm\-roberta\-large666[https://huggingface\.co/FacebookAI/xlm\-roberta\-large](https://huggingface.co/FacebookAI/xlm-roberta-large)\([Conneau et al\., 2020](https://arxiv.org/html/2610.11520#bib.bib35)\)as text encoder for all three languages\. It has French as high\-resource language \(9\.8 billion tokens\), Slovenian as medium\-resource language \(1\.7 billion tokens\), and Hausa as low\-resource language \(56 million tokens\) in its pre\-training data\.

We use the same hyperparameters across all architectures and datasets \(see Table[5](https://arxiv.org/html/2610.11520#A1.T5)in Appendix[A](https://arxiv.org/html/2610.11520#A1)for the full list\)\. We select the best checkpoint on the development set in each setting, and report the results on the test set\. Our implementation uses the speechbrain library\([Ravanelli et al\., 2021](https://arxiv.org/html/2610.11520#bib.bib29)\)\.

Evaluation metricsWe use standard evaluation measures: Word Error Rate \(WER\) for speech recognition, POS accuracy \(POS\), Unlabeled Attachment Score \(UAS\), and Labeled Attachment Score \(LAS\) for dependency parsing\. The dependency parsing measures require an alignment between the ground truth and the predicted sequences, which was performed following the procedure described by[Pupier et al\. \(2024\)](https://arxiv.org/html/2610.11520#bib.bib19)\.

## 6Results

### 6\.1RQ1: Role of Intermediate NN Units

##### \(H1\) Simplified architecture

In Table[2](https://arxiv.org/html/2610.11520#S5.T2), we show the comparison of baseline and simplified Wav2tree across all three spoken treebanks \(1stand 2ndrows of each sub\-table\)\.

After removing the three intermediate NN units, parsing accuracy \(LAS\) improved for Slovenian and French \(Paris Stories\)\. In absolute terms, LAS increased by 1\.2%, and 2\.5% on Slovenian and French\. There is a slight drop for Naija \(LAS decreased by 0\.95%\)\. For Slovenian and French, the LAS improvement from the Baseline to the Simplified architecture is statistically significant \(see Appendix[D](https://arxiv.org/html/2610.11520#A4)for details\)\. For all three datasets, ASR performance remains consistent across the Baseline and Simplified Wav2tree\.

Based on this evaluation, we infer that the Simplified W2T, while being parameter\-efficient \(see Appendix[E](https://arxiv.org/html/2610.11520#A5)for the effect on training time\), matches or improves parsing performance compared to Wav2tree\. Our results support hypothesis H1 that intermediate NN units do not provide incremental value when the speech model is jointly fine\-tuned with the parsing model\.

##### \(H2\) When do intermediate NN units add value?

In Table[3](https://arxiv.org/html/2610.11520#S5.T3), we report the text parsing performance \(LAS score\) on all three treebanks across the combinations of \(i\) text parsing architectures with and without an intermediate NN unit \(LSTM\), and \(ii\) whether the pre\-trained text model is frozen or fine\-tuned jointly with the parsing model\.

Across all three treebanks, when the pre\-trained text model is fine\-tuned jointly with the parsing model \(1stand 2ndrows\), LAS score is higher when there is no intermediate NN unit between the text encoder and the parsing model\. This is consistent with our findings from the Simplified architecture\.

When we freeze the pre\-trained text model and feed its output directly to the parsing model \(3rdrow\), the LAS score drops by at least 30% across all three treebanks\. We attribute this to therepresentational gapbetween the frozen encoder and the parsing model\. We then introduce an intermediate NN unit into this setup \(4throw\), and the LAS score increases by 15\-20% across all three treebanks, bringing the performance closer to the joint fine\-tuning setting\. This suggests that intermediate NN unit helps to reduce therepresentational gapbetween the frozen pre\-trained model and the parsing model\.

We then measure the similarity between the text representation from the fine\-tuned text model without any intermediate NN unit and \(i\) the representation from the frozen text model without any intermediate NN unit\. The representation corresponds to the encoder output of the pre\-trained model \(ii\) the representation from the frozen text model with an intermediate NN unit\. The representation corresponds to the output of the intermediate NN unit\.

We use linear Centered Kernel Alignment[Kornblith et al\. \(2019\)](https://arxiv.org/html/2610.11520#bib.bib38)as the similarity measure\. Centered Kernel Alignment \(CKA\) is a widely used pairwise feature similarity for quantifying the similarity between two neural network representations\. It measures the similarity between features from the two representations for each pair of examples using Hilbert\-Schmidt Independence Criterion\.

The similarity is shown in Table[4](https://arxiv.org/html/2610.11520#S6.T4)\. Across all three treebanks, the similarity between the representations from the fine\-tuned and the frozen text model is higher with an intermediate NN unit than without\.

Based on these two evaluations, we answer RQ1 and support the hypothesis H2 that the intermediate NN unit helps to reduce therepresentational gapbetween the frozen pre\-trained model and the parsing model\.

### 6\.2RQ2: Scaling Training Data

In Figure[4](https://arxiv.org/html/2610.11520#S6.F4), we report the performance of Naija and French \(Orféo\) across the two setups described in Section[5\.2](https://arxiv.org/html/2610.11520#S5.SS2)\. The WER is higher than usual for Orféo because of the nature of the speech corpus \(spontaneous discussions, and lower recording quality for some sub\-corpora\)\.

Across all treebanks, in the first experiment \(Baseline setup\), the WER decreases and the LAS improves with a larger training set\. In the subsequent experiment \(Frozen ASR\), where a pre\-fine\-tuned ASR is used across all data partitions, the LAS increases as the training data scales up, the magnitude of gains is reduced compared to the first experiment\. Here, the LAS trend differs across the two spoken treebanks – the LAS remains relatively stable across all data partitions for Orféo, while it increases up to 80% of the training data for Naija\. We attribute this to the difference in the amount of the training data of the two treebanks\.

This analysis suggests that with a pre\-fine\-tuned ASR model, parsing accuracy gains diminish beyond 9\-10 hours of training data with the Simplified W2T\.

### 6\.3Dependency Parsing Evaluation

In Table[2](https://arxiv.org/html/2610.11520#S5.T2), we report the evaluation for Wav2Vec2 intermediate layers, and the baseline results for all three spoken treebanks\.

##### Intermediate layers

The 3rdrow in Table[2](https://arxiv.org/html/2610.11520#S5.T2)corresponds to the performance when the intermediate layer of the Wav2Vec2 model is used to produce the speech representation, as described in Section[5\.3](https://arxiv.org/html/2610.11520#S5.SS3)\. We report results \(on the test dataset\) for the layer which has highest LAS score on the development dataset\. Across all three spoken treebanks, the LAS score remains stable or improved\. This corresponds to a 12%\-24% reduction in the parameters and 5\-9% speed up in training time\. We conclude that intermediate layers of the speech encoder produce speech representations suitable for parsing\.

##### Pipeline and GOLD

The last two rows in Table[2](https://arxiv.org/html/2610.11520#S5.T2)correspond to the Pipeline and the GOLD setups\. Apart from the Slovenian treebank, the end\-to\-end Simplified architecture performs better than the Pipeline approach on parsing, which also has 2\.75×\\timesmore parameters than the Simplified architecture, due to the use of two pre\-trained models: an acoustic model for ASR and a pretrained text model for parsing\.

##### WER for Naija

We compare word error rate \(WER\) for tokens in Naija depending on whether they are in the UnixwordsEnglish lexicon\.777[https://www\.nltk\.org/book/ch02\.html](https://www.nltk.org/book/ch02.html)We find that WER is substantially higher for tokens not belonging to the English vocabulary, as displayed in Table[7](https://arxiv.org/html/2610.11520#A1.T7)in Appendix\. This result is likely due to the limited size of Naija\-related languages in its pretraining data, and illustrates the English bias ofwav2vec2\-xls\-r\-300m\.

Similarity betweenCKACKACKAfine\-tuned encoder &NaijaSlovenianFrenchFrozen encoderwithout17%21%23%intermediate NN unitFrozen encoderwith44%32%34%intermediate NN unit

Table 4:Similarity between representations from fine\-tuned and frozen text encoder described in Section[6\.1](https://arxiv.org/html/2610.11520#S6.SS1.SSS0.Px2)![Refer to caption](https://arxiv.org/html/2610.11520v1/figs/ScalingNaija.png)

\(\(a\)\)Data Scaling \- Naija
![Refer to caption](https://arxiv.org/html/2610.11520v1/figs/ScalingOrfeo.png)

\(\(b\)\)Data Scaling \- Orféo

Figure 4:Scaling training data across 2 setups \- End\-to\-End \(Baseline\) and Frozen ASR \- as described in Section[5\.2](https://arxiv.org/html/2610.11520#S5.SS2)

## 7Conclusion and Further work

We investigate the role of intermediate neural network units in the parsing architecture and find that they provide incremental value when the pre\-trained model is frozen by reducing therepresentational gapbetween the pre\-trained model and parsing model\. We further introduce an end\-to\-end parser for speech that is parameter\-efficient while achieving comparable performance to the prior architecture\. We evaluate the impact of scaling training data on the Simplified architecture, and find that gains in parsing accuracy diminish after the first 9\-10h of training data\.

We present baseline dependency parsing results for three spoken treebanks — Naija, Slovenian and French \(ParisStories\)\. We provide the first end\-to\-end speech parsing experiments for Naija and Slovenian\. We assess the impact of speech representation from intermediate layers of Wav2Vec2 as the input to the parsing architecture, and find that the parsing performance remains stable\. Lastly, we release the code to reproduce filteredUD\_Slovenian\-SSTv2\.16 treebank with improved speech\-transcription alignment

The analysis of intermediate neural network units can be extended to other neural architectures beyond parsing\. As the future work we plan to examine two findings from this paper: \(i\) why the pipeline performs better than the end\-to\-end speech parser on the Slovenian treebank, and \(ii\) why parsing performance remains stable when using representations from the intermediate layer of the speech encoder as the input\.

## Limitations

For all our experiments, we limited our hyperparameter search to learning rate due to computational resource constraints\. We do not evaluate the proposed architecture on the English Switchboard corpus\([Godfrey et al\., 1992](https://arxiv.org/html/2610.11520#bib.bib7)\)because the audio recordings are not publicly available\.

## Acknowledgements

We gratefully acknowledge the support of the French National Research Agency \(ANR\) through project SynPaX \(grant ANR\-23\-CE23\-0017\-01\)\. We thank William Havard for feedback on an earlier version of this paper, as well as Rayan Ziane and Emmanuel Schang for fruitful discussions about this work\.

## References

- Baevskiet al\.\(2022\)A\. Baevski, W\. Hsu, Q\. Xu, A\. Babu, J\. Gu, and M\. AuliData2vec: a general framework for self\-supervised learning in speech, vision and language\.InProceedings of the 39th International Conference on Machine Learning,K\. Chaudhuri, S\. Jegelka, L\. Song, C\. Szepesvari, G\. Niu, and S\. Sabato \(Eds\.\),Proceedings of Machine Learning Research, Vol\.162,pp\. 1298–1312\.External Links:[Link](https://proceedings.mlr.press/v162/baevski22a.html)Cited by:[§5\.5](https://arxiv.org/html/2610.11520#S5.SS5.SSS0.Px1.p1.1)\.
- Baevskiet al\.\(2020\)A\. Baevski, Y\. Zhou, A\. Mohamed, and M\. AuliWav2vec 2\.0: a framework for self\-supervised learning of speech representations\.Advances in neural information processing systems33,pp\. 12449–12460\.Cited by:[2nd item](https://arxiv.org/html/2610.11520#S1.I1.i2.I1.i2.p1.1)\.
- Béchetet al\.\(2014\)F\. Béchet, A\. Nasr, and B\. FavreAdapting dependency parsing to spontaneous speech for open domain spoken language understanding\.InInterspeech 2014,pp\. 135–139\.External Links:[Document](https://dx.doi.org/10.21437/Interspeech.2014-39),ISSN 2958\-1796Cited by:[§2](https://arxiv.org/html/2610.11520#S2.p1.1)\.
- Benzitounet al\.\(2016\)C\. Benzitoun, J\. Debaisieux, and H\. DeulofeuLe projet ORFÉO: un corpus d’étude pour le français contemporain\.Corpus\(15\)\.Cited by:[§2](https://arxiv.org/html/2610.11520#S2.p2.1),[§3](https://arxiv.org/html/2610.11520#S3.p1.1)\.
- Caronet al\.\(2019\)B\. Caron, M\. Courtin, K\. Gerdes, and S\. KahaneA surface\-syntactic UD treebank for Naija\.InProceedings of the 18th International Workshop on Treebanks and Linguistic Theories \(TLT, SyntaxFest 2019\),M\. Candito, K\. Evang, S\. Oepen, and D\. Seddah \(Eds\.\),Paris, France,pp\. 13–24\.External Links:[Link](https://aclanthology.org/W19-7803/),[Document](https://dx.doi.org/10.18653/v1/W19-7803)Cited by:[4th item](https://arxiv.org/html/2610.11520#A2.I1.i4.p1.1),[Appendix B](https://arxiv.org/html/2610.11520#A2.p3.1),[Appendix B](https://arxiv.org/html/2610.11520#A2.p5.1),[item 1](https://arxiv.org/html/2610.11520#S3.I1.i1.p1.1),[§5\.4](https://arxiv.org/html/2610.11520#S5.SS4.p1.1),[footnote 8](https://arxiv.org/html/2610.11520#footnote8)\.
- Charniak and Johnson \(2001\)E\. Charniak and M\. JohnsonEdit detection and parsing for transcribed speech\.InSecond Meeting of the North American Chapter of the Association for Computational Linguistics,External Links:[Link](https://aclanthology.org/N01-1016)Cited by:[§2](https://arxiv.org/html/2610.11520#S2.p1.1)\.
- Chrupała \(2023\)G\. ChrupałaPutting natural in natural language processing\.InFindings of the Association for Computational Linguistics: ACL 2023,A\. Rogers, J\. Boyd\-Graber, and N\. Okazaki \(Eds\.\),Toronto, Canada,pp\. 7820–7827\.External Links:[Link](https://aclanthology.org/2023.findings-acl.495/),[Document](https://dx.doi.org/10.18653/v1/2023.findings-acl.495)Cited by:[§1](https://arxiv.org/html/2610.11520#S1.p2.1)\.
- Conneauet al\.\(2020\)A\. Conneau, K\. Khandelwal, N\. Goyal, V\. Chaudhary, G\. Wenzek, F\. Guzmán, E\. Grave, M\. Ott, L\. Zettlemoyer, and V\. StoyanovUnsupervised cross\-lingual representation learning at scale\.External Links:1911\.02116,[Link](https://arxiv.org/abs/1911.02116)Cited by:[§5\.5](https://arxiv.org/html/2610.11520#S5.SS5.SSS0.Px2.p1.1)\.
- Demšar \(2006\)J\. DemšarStatistical comparisons of classifiers over multiple data sets\.Journal of Machine learning research7\(Jan\),pp\. 1–30\.Cited by:[§D\.1](https://arxiv.org/html/2610.11520#A4.SS1.p1.1)\.
- Dobrovoljc and Nivre \(2016\)K\. Dobrovoljc and J\. NivreThe Universal Dependencies treebank of spoken Slovenian\.InProceedings of the Tenth International Conference on Language Resources and Evaluation \(LREC’16\),N\. Calzolari, K\. Choukri, T\. Declerck, S\. Goggi, M\. Grobelnik, B\. Maegaard, J\. Mariani, H\. Mazo, A\. Moreno, J\. Odijk, and S\. Piperidis \(Eds\.\),Portorož, Slovenia,pp\. 1566–1573\.External Links:[Link](https://aclanthology.org/L16-1248/)Cited by:[item 2](https://arxiv.org/html/2610.11520#S3.I1.i2.p1.1)\.
- Dozat and Manning \(2017\)T\. Dozat and C\. D\. ManningDeep biaffine attention for neural dependency parsing\.In5th International Conference on Learning Representations, ICLR 2017, Toulon, France, April 24\-26, 2017, Conference Track Proceedings,External Links:[Link](https://openreview.net/forum?id=Hk95PK9le)Cited by:[§4](https://arxiv.org/html/2610.11520#S4.p5.3),[§5\.5](https://arxiv.org/html/2610.11520#S5.SS5.SSS0.Px1.p2.1)\.
- Dreyer and Shafran \(2007\)M\. Dreyer and I\. ShafranExploiting prosody for PCFGs with latent annotations\.InInterspeech 2007,pp\. 450–453\.External Links:[Document](https://dx.doi.org/10.21437/Interspeech.2007-216),ISSN 2958\-1796Cited by:[§2](https://arxiv.org/html/2610.11520#S2.p1.1)\.
- Dugonjićet al\.\(2024\)Z\. Dugonjić, A\. Pupier, B\. Lecouteux, and M\. CoavouxWhat has LeBenchmark learnt about French syntax?\.InProceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation \(LREC\-COLING 2024\),N\. Calzolari, M\. Kan, V\. Hoste, A\. Lenci, S\. Sakti, and N\. Xue \(Eds\.\),Torino, Italia,pp\. 17493–17499\.External Links:[Link](https://aclanthology.org/2024.lrec-main.1521/)Cited by:[§5\.3](https://arxiv.org/html/2610.11520#S5.SS3.SSS0.Px1.p1.1)\.
- Gerdeset al\.\(2018\)K\. Gerdes, B\. Guillaume, S\. Kahane, and G\. PerrierSUD or surface\-syntactic Universal Dependencies: an annotation scheme near\-isomorphic to UD\.InProceedings of the Second Workshop on Universal Dependencies \(UDW 2018\),M\. de Marneffe, T\. Lynn, and S\. Schuster \(Eds\.\),Brussels, Belgium,pp\. 66–74\.External Links:[Link](https://aclanthology.org/W18-6008/),[Document](https://dx.doi.org/10.18653/v1/W18-6008)Cited by:[item 1](https://arxiv.org/html/2610.11520#S3.I1.i1.p1.1)\.
- Godfreyet al\.\(1992\)J\. J\. Godfrey, E\. C\. Holliman, and J\. McDanielSWITCHBOARD: telephone speech corpus for research and development\.In\[Proceedings\] ICASSP\-92: 1992 IEEE International Conference on Acoustics, Speech, and Signal Processing,Vol\.1,pp\. 517–520 vol\.1\.External Links:[Document](https://dx.doi.org/10.1109/ICASSP.1992.225858)Cited by:[§2](https://arxiv.org/html/2610.11520#S2.p2.1),[Limitations](https://arxiv.org/html/2610.11520#Sx1.p1.1)\.
- Graveset al\.\(2006\)A\. Graves, S\. Fernández, F\. Gomez, and J\. SchmidhuberConnectionist temporal classification: labelling unsegmented sequence data with recurrent neural networks\.InProceedings of the 23rd international conference on Machine learning,pp\. 369–376\.Cited by:[§4](https://arxiv.org/html/2610.11520#S4.p2.2)\.
- Honnibal and Johnson \(2014\)M\. Honnibal and M\. JohnsonJoint incremental disfluency detection and dependency parsing\.Transactions of the Association for Computational Linguistics2,pp\. 131–142\.External Links:[Link](https://aclanthology.org/Q14-1011),[Document](https://dx.doi.org/10.1162/tacl%5Fa%5F00171)Cited by:[§2](https://arxiv.org/html/2610.11520#S2.p1.1)\.
- Hüll and Dobrovoljc \(2025\)N\. Hüll and K\. DobrovoljcWord order variation in spoken and written corpora: a cross\-linguistic study of SVO and alternative orders\.InProceedings of the Eighth International Conference on Dependency Linguistics \(Depling, SyntaxFest 2025\),E\. Hajičová and S\. Kahane \(Eds\.\),Ljubljana, Slovenia,pp\. 150–155\.External Links:[Link](https://aclanthology.org/2025.depling-1.16/),ISBN 979\-8\-89176\-290\-9Cited by:[§1](https://arxiv.org/html/2610.11520#S1.p2.1)\.
- Jamshid Louet al\.\(2019\)P\. Jamshid Lou, Y\. Wang, and M\. JohnsonNeural constituency parsing of speech transcripts\.InProceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 \(Long and Short Papers\),J\. Burstein, C\. Doran, and T\. Solorio \(Eds\.\),Minneapolis, Minnesota,pp\. 2756–2765\.External Links:[Link](https://aclanthology.org/N19-1282/),[Document](https://dx.doi.org/10.18653/v1/N19-1282)Cited by:[§2](https://arxiv.org/html/2610.11520#S2.p1.1)\.
- Jørgensen \(2007\)F\. JørgensenThe effects of disfluency detection in parsing spoken language\.InProceedings of the 16th Nordic Conference of Computational Linguistics \(NODALIDA 2007\),J\. Nivre, H\. Kaalep, K\. Muischnek, and M\. Koit \(Eds\.\),Tartu, Estonia,pp\. 240–244\.External Links:[Link](https://aclanthology.org/W07-2435)Cited by:[§2](https://arxiv.org/html/2610.11520#S2.p1.1)\.
- Kahnet al\.\(2005\)J\. G\. Kahn, M\. Lease, E\. Charniak, M\. Johnson, and M\. OstendorfEffective use of prosody in parsing conversational speech\.InProceedings of Human Language Technology Conference and Conference on Empirical Methods in Natural Language Processing,R\. Mooney, C\. Brew, L\. Chien, and K\. Kirchhoff \(Eds\.\),Vancouver, British Columbia, Canada,pp\. 233–240\.External Links:[Link](https://aclanthology.org/H05-1030)Cited by:[§2](https://arxiv.org/html/2610.11520#S2.p1.1)\.
- Kandoet al\.\(2024\)S\. Kando, Y\. Miyao, J\. Naradowsky, and S\. TakamichiTextless dependency parsing by labeled sequence prediction\.InInterspeech 2024,pp\. 1340–1344\.External Links:[Document](https://dx.doi.org/10.21437/Interspeech.2024-783),ISSN 2958\-1796Cited by:[§1](https://arxiv.org/html/2610.11520#S1.p1.1),[§2](https://arxiv.org/html/2610.11520#S2.p2.1)\.
- Kornblithet al\.\(2019\)S\. Kornblith, M\. Norouzi, H\. Lee, and G\. HintonSimilarity of neural network representations revisited\.InInternational conference on machine learning,pp\. 3519–3529\.Cited by:[§6\.1](https://arxiv.org/html/2610.11520#S6.SS1.SSS0.Px2.p5.1)\.
- Laiet al\.\(2023\)C\. J\. Lai, F\. Shi, P\. Peng, Y\. Kim, K\. Gimpel, S\. Chang, Y\. Chuang, S\. Bhati, D\. Cox, D\. Harwath, Y\. Zhang, K\. Livescu, and J\. GlassAudio\-visual neural syntax acquisition\.External Links:2310\.07654,[Link](https://arxiv.org/abs/2310.07654)Cited by:[§2](https://arxiv.org/html/2610.11520#S2.p3.1)\.
- Leet al\.\(2026\)P\. Le, V\. Pelloin, A\. Chatelain, M\. Bouziane, M\. Ghennai, Q\. Guan, K\. Milintsevich, S\. Mdhaffar, A\. Mannion, N\. Defauw, S\. Gu, A\. D\. Audibert, M\. Dinarelli, Y\. Estève, L\. Goeuriot, S\. Lalande, N\. Hervé, M\. Coavoux, F\. Portet, É\. Ollion, M\. Candito, M\. Peyrard, S\. Rossato, B\. Lecouteux, A\. Nardy, G\. Sérasset, V\. Segonne, S\. Evain, D\. Fabre, and D\. SchwabPantagruel: unified self\-supervised encoders for french text and speech\.InProceedings of the Fifteenth Language Resources and Evaluation Conference \(LREC 2026\),S\. Piperidis, N\. Bel, H\. van den Heuvel, N\. Ide, S\. Krek, and A\. Toral \(Eds\.\),Palma, Mallorca, Spain,pp\. 10168–10191\.External Links:[Document](https://dx.doi.org/10.63317/573q4exhmpgd)Cited by:[§5\.5](https://arxiv.org/html/2610.11520#S5.SS5.SSS0.Px1.p1.1)\.
- Linget al\.\(2015\)W\. Ling, C\. Dyer, A\. W\. Black, I\. Trancoso, R\. Fermandez, S\. Amir, L\. Marujo, and T\. LuísFinding function in form: compositional character models for open vocabulary word representation\.InProceedings of the 2015 Conference on Empirical Methods in Natural Language Processing,L\. Màrquez, C\. Callison\-Burch, and J\. Su \(Eds\.\),Lisbon, Portugal,pp\. 1520–1530\.External Links:[Link](https://aclanthology.org/D15-1176/),[Document](https://dx.doi.org/10.18653/v1/D15-1176)Cited by:[§4](https://arxiv.org/html/2610.11520#S4.p5.1)\.
- Nivreet al\.\(2020\)J\. Nivre, M\. de Marneffe, F\. Ginter, J\. Hajič, C\. D\. Manning, S\. Pyysalo, S\. Schuster, F\. Tyers, and D\. ZemanUniversal Dependencies v2: an evergrowing multilingual treebank collection\.InProceedings of the Twelfth Language Resources and Evaluation Conference,N\. Calzolari, F\. Béchet, P\. Blache, K\. Choukri, C\. Cieri, T\. Declerck, S\. Goggi, H\. Isahara, B\. Maegaard, J\. Mariani, H\. Mazo, A\. Moreno, J\. Odijk, and S\. Piperidis \(Eds\.\),Marseille, France,pp\. 4034–4043\(eng\)\.External Links:[Link](https://aclanthology.org/2020.lrec-1.497/),ISBN 979\-10\-95546\-34\-4Cited by:[item 1](https://arxiv.org/html/2610.11520#S3.I1.i1.p1.1),[§3](https://arxiv.org/html/2610.11520#S3.p1.1)\.
- Pate and Goldwater \(2013\)J\. K\. Pate and S\. GoldwaterUnsupervised dependency parsing with acoustic cues\.Transactions of the Association for Computational Linguistics1,pp\. 63–74\.External Links:[Link](https://aclanthology.org/Q13-1006),[Document](https://dx.doi.org/10.1162/tacl%5Fa%5F00210)Cited by:[§2](https://arxiv.org/html/2610.11520#S2.p1.1)\.
- Peck and Becker \(2024\)N\. Peck and L\. BeckerSyntactic pausing? re\-examining the associations\.Linguistics Vanguard10\(1\),pp\. 223–237\.External Links:[Link](https://doi.org/10.1515/lingvan-2022-0156),[Document](https://dx.doi.org/doi%3A10.1515/lingvan-2022-0156)Cited by:[§1](https://arxiv.org/html/2610.11520#S1.p2.1)\.
- Planket al\.\(2016\)B\. Plank, A\. Søgaard, and Y\. GoldbergMultilingual part\-of\-speech tagging with bidirectional long short\-term memory models and auxiliary loss\.InProceedings of the 54th Annual Meeting of the Association for Computational Linguistics \(Volume 2: Short Papers\),K\. Erk and N\. A\. Smith \(Eds\.\),Berlin, Germany,pp\. 412–418\.External Links:[Link](https://aclanthology.org/P16-2067/),[Document](https://dx.doi.org/10.18653/v1/P16-2067)Cited by:[§4](https://arxiv.org/html/2610.11520#S4.p5.1)\.
- Pratapet al\.\(2023\)V\. Pratap, A\. Tjandra, B\. Shi, P\. Tomasello, A\. Babu, S\. Kundu, A\. Elkahky, Z\. Ni, A\. Vyas, M\. Fazel\-Zarandi, A\. Baevski, Y\. Adi, X\. Zhang, W\. Hsu, A\. Conneau, and M\. AuliScaling speech technology to 1,000\+ languages\.External Links:2305\.13516,[Link](https://arxiv.org/abs/2305.13516)Cited by:[Appendix C](https://arxiv.org/html/2610.11520#A3.p2.1)\.
- Pupieret al\.\(2024\)A\. Pupier, M\. Coavoux, J\. Goulian, and B\. LecouteuxGrowing trees on sounds: assessing strategies for end\-to\-end dependency parsing of speech\.InProceedings of the 62nd Annual Meeting of the Association for Computational Linguistics \(Volume 2: Short Papers\),L\. Ku, A\. Martins, and V\. Srikumar \(Eds\.\),Bangkok, Thailand,pp\. 225–233\.External Links:[Link](https://aclanthology.org/2024.acl-short.22/),[Document](https://dx.doi.org/10.18653/v1/2024.acl-short.22)Cited by:[Table 5](https://arxiv.org/html/2610.11520#A1.T5),[Table 5](https://arxiv.org/html/2610.11520#A1.T5.2.1.1.2.1),[Appendix E](https://arxiv.org/html/2610.11520#A5.p2.1),[§2](https://arxiv.org/html/2610.11520#S2.p2.1),[Figure 2](https://arxiv.org/html/2610.11520#S3.F2),[2\(a\)](https://arxiv.org/html/2610.11520#S3.F2.sf1),[§5\.1\.1](https://arxiv.org/html/2610.11520#S5.SS1.SSS1.p5.1),[§5\.5](https://arxiv.org/html/2610.11520#S5.SS5.SSS0.Px2.p3.1)\.
- Pupieret al\.\(2022\)A\. Pupier, M\. Coavoux, B\. Lecouteux, and J\. GoulianEnd\-to\-end dependency parsing of spoken French\.InInterspeech 2022,pp\. 1816–1820\.External Links:[Document](https://dx.doi.org/10.21437/Interspeech.2022-381),ISSN 2958\-1796Cited by:[§1](https://arxiv.org/html/2610.11520#S1.p1.1),[§1](https://arxiv.org/html/2610.11520#S1.p3.1),[§2](https://arxiv.org/html/2610.11520#S2.p1.1),[§2](https://arxiv.org/html/2610.11520#S2.p2.1),[§4](https://arxiv.org/html/2610.11520#S4.p1.1)\.
- Rasooli and Tetreault \(2013\)M\. S\. Rasooli and J\. TetreaultJoint parsing and disfluency detection in linear time\.InProceedings of the 2013 Conference on Empirical Methods in Natural Language Processing,D\. Yarowsky, T\. Baldwin, A\. Korhonen, K\. Livescu, and S\. Bethard \(Eds\.\),Seattle, Washington, USA,pp\. 124–129\.External Links:[Link](https://aclanthology.org/D13-1013)Cited by:[§2](https://arxiv.org/html/2610.11520#S2.p1.1)\.
- Rasooli and Tetreault \(2014\)M\. S\. Rasooli and J\. TetreaultNon\-monotonic parsing of fluent umm I mean disfluent sentences\.InProceedings of the 14th Conference of the European Chapter of the Association for Computational Linguistics, volume 2: Short Papers,S\. Wintner, S\. Riezler, and S\. Goldwater \(Eds\.\),Gothenburg, Sweden,pp\. 48–53\.External Links:[Link](https://aclanthology.org/E14-4010/),[Document](https://dx.doi.org/10.3115/v1/E14-4010)Cited by:[§2](https://arxiv.org/html/2610.11520#S2.p1.1)\.
- Ravanelliet al\.\(2021\)M\. Ravanelli, T\. Parcollet, P\. Plantinga, A\. Rouhe, S\. Cornell, L\. Lugosch, C\. Subakan, N\. Dawalatabad, A\. Heba, J\. Zhong, J\. Chou, S\. Yeh, S\. Fu, C\. Liao, E\. Rastorgueva, F\. Grondin, W\. Aris, H\. Na, Y\. Gao, R\. D\. Mori, and Y\. BengioSpeechBrain: a general\-purpose speech toolkit\.External Links:2106\.04624,[Link](https://arxiv.org/abs/2106.04624)Cited by:[§4](https://arxiv.org/html/2610.11520#S4.p3.1),[§5\.5](https://arxiv.org/html/2610.11520#S5.SS5.SSS0.Px2.p2.1)\.
- Shenet al\.\(2023\)G\. Shen, A\. Alishahi, A\. Bisazza, and G\. ChrupałaWave to syntax: probing spoken language models for syntax\.InInterspeech 2023,pp\. 1259–1263\.External Links:[Document](https://dx.doi.org/10.21437/Interspeech.2023-679),ISSN 2958\-1796Cited by:[§5\.3](https://arxiv.org/html/2610.11520#S5.SS3.SSS0.Px1.p1.1)\.
- Strzyzet al\.\(2019\)M\. Strzyz, D\. Vilares, and C\. Gómez\-RodríguezViable dependency parsing as sequence labeling\.InNAACL 2019,External Links:[Link](https://aclanthology.org/N19-1077)Cited by:[§4](https://arxiv.org/html/2610.11520#S4.p5.3)\.
- Tran and Ostendorf \(2021\)T\. Tran and M\. OstendorfAssessing the use of prosody in constituency parsing of imperfect transcripts\.InProc\. Interspeech 2021,pp\. 2626–2630\.External Links:[Document](https://dx.doi.org/10.21437/Interspeech.2021-373)Cited by:[§2](https://arxiv.org/html/2610.11520#S2.p1.1)\.
- Tsenget al\.\(2023\)Y\. Tseng, C\. J\. Lai, and H\. LeeCascading and direct approaches to unsupervised constituency parsing on spoken sentences\.InICASSP 2023 \- 2023 IEEE International Conference on Acoustics, Speech and Signal Processing \(ICASSP\),Vol\.,pp\. 1–5\.External Links:[Document](https://dx.doi.org/10.1109/ICASSP49357.2023.10094575)Cited by:[§2](https://arxiv.org/html/2610.11520#S2.p3.1)\.
- Yoshikawaet al\.\(2016\)M\. Yoshikawa, H\. Shindo, and Y\. MatsumotoJoint transition\-based dependency parsing and disfluency detection for automatic speech recognition texts\.InProceedings of the 2016 Conference on Empirical Methods in Natural Language Processing,J\. Su, K\. Duh, and X\. Carreras \(Eds\.\),Austin, Texas,pp\. 1036–1041\.External Links:[Link](https://aclanthology.org/D16-1109),[Document](https://dx.doi.org/10.18653/v1/D16-1109)Cited by:[§2](https://arxiv.org/html/2610.11520#S2.p1.1)\.

## Appendix AHyperparameters

We provide the full list of hyperparameters for our experiments in Table[5](https://arxiv.org/html/2610.11520#A1.T5)\.

ArchitectureBaseline Wav2tree\([Pupier et al\., 2024](https://arxiv.org/html/2610.11520#bib.bib19)\)Simplified Wav2tree \(ours\)Epoch8080Batch size44Tuning parametersLearning rate \(ASR\)0\.000050\.00005Learning rate \(Remaining\)0\.00030\.0003OptimizerAdamWAdamWModel namewav2vec2\-xls\-r\-300m \(Naija & Slovenian\)PantagrueLLM/Speech\_Large\_fr\_114K \(ParisStories\)wav2vec2\-xls\-r\-300m \(Naija & Slovenian\)PantagrueLLM/Speech\_Large\_fr\_114K \(ParisStories\)EncoderPresentN/AEncoder layer3Dropout0\.15Encoder dim1024ActivationLeakyReLUFusion LSTMPresentN/ALayer2Dim500BidirectionalFALSEBiasTRUELSTM parserPresentN/ALayer3Dim256BidirectionalTRUEArc MLPDim256256Layer11Linear head dim256256Label MLPDim256256Layer11Head dim256256POS MLPDim256256Linear head dimDependent on the dataDependent on the data

Table 5:Hyperparameters across Wav2tree architecture\([Pupier et al\., 2024](https://arxiv.org/html/2610.11520#bib.bib19)\)and Simplified Wav2tree \(ours\)\.![Refer to caption](https://arxiv.org/html/2610.11520v1/figs/CTCexample.png)

Figure 5:Predicted word boundary and the corresponding uncollapsed sequence of the predicted transcription \(in French\) "J AI MON COLLÈGUE" \(en:I have my colleague\)\. Symbolsϵ\\epsilonand \_ represents respectively the blank and space labels\. Indices \{1,2,3,4\} refer to the word position in the output sequence\.IndexPunctuation presenceWER ↓UPOS ↑UAS ↑LAS ↑1Train & Eval data34\.476\.562\.857\.62Training data3278\.568\.861\.83None31\.879\.169\.161\.8

Table 6:Simplified W2T results across different variations of Naija dataset\. Best scores highlighted inbold\.![Refer to caption](https://arxiv.org/html/2610.11520v1/figs/WAZL_15_MC-Abi_MG__180_gold_annotation.png)

\(\(a\)\)Gold annotation
![Refer to caption](https://arxiv.org/html/2610.11520v1/figs/WAZL_15_MC-Abi_MG__180_predicted_annotation.png)

\(\(b\)\)Predicted annotation

Figure 6:Gold and predicted transcription and trees for a Naija utterance \(sent\_id: WAZL\_15\_MC\_Abi\-MG\_\_180\)\. The predicted annotation has all UAS, UPOS and LAS as 72\.7\. The scores are low because not all markups are predicted by ASR \- ‘\(’, ‘\)’ and ‘<<’\(marked as ‘&lt’\) are not predicted by ASR\.English VocabNumber of wordsWER18,99124\.5%01,26952\.5%

Table 7:WER for tokens in Naija\-NSC treebank\.
## Appendix BPunctuation in Naija

The main punctuation marks in Naija dataset are:888We refer the reader to[Caron et al\. \(2019\)](https://arxiv.org/html/2610.11520#bib.bib1)for a more thorough description and examples\.

- •//delimits illocutionary units \(final token of each dependency tree\)
- •a bracketing with\{ \}indicates a sequence of elements with the same syntactic function; elements in the sequence are delimited by a separator that depends on the type of the sequence\. For example, in Figure[1](https://arxiv.org/html/2610.11520#S1.F1), a disfluency with a false start is represented with the pattern\{ reparandum \|\| repair \}\. Possible separators are\|\|for reformulations and disfluencies,\|cfor coordination,\|rfor syntactic reduplication and\|afor apposition\.
- •\#denotes a short pause\.
- •<and\>separate the central part of the utterance \(nucleus\) from the peripheral elements\(ad\-nuclei, see[Caron et al\., 2019](https://arxiv.org/html/2610.11520#bib.bib1), and Figure[6\(a\)](https://arxiv.org/html/2610.11520#A1.F6.sf1)\)\.

We created three setups to investigate the impact of the punctuation marks for speech parsing\.

1. 1\.Both training and evaluation datasets have markups\.
2. 2\.Only training dataset has markups\. During evaluation, markups are removed from the prediction before computing ASR and parsing measures
3. 3\.Neither training nor evaluation datasets have markups

We identify punctuation marks as characters that do not contain alphabetic characters or numbers\. We first assess the utility of punctuation marks in the evaluation dataset \(comparing the first 2 setups\)\. This evaluation is aligned with the the comparison setting of[Caron et al\. \(2019\)](https://arxiv.org/html/2610.11520#bib.bib1), who evaluated the parsing performance on treebank with and without markups\. Table[6](https://arxiv.org/html/2610.11520#A1.T6)shows this comparison \(1stand 2ndrows\)\. The LAS score drops by 4\.1% \(in absolute terms\) if there are punctuation marks in the evaluation dataset\.

We dig deeper into why the presence of segmentation annotations in the transcription have a negative impact on parsing\. We find a correlation between the ASR punctuation recognition errors and parsing accuracy\. Specifically, as the number of markups increases in the transcript, not all these markups are predicted by ASR, which impacts the performance of both ASR and parsing\. Table[8](https://arxiv.org/html/2610.11520#A2.T8)displays the performance of evaluation data with and without markups, stratified by the number of markups present in the ground truth transcript\. As the number of markups in the transcript increases, ASR predicts fewer markups than present in the ground truth, and this degradation also affects LAS, as evidenced by the increasing widening gap between LAS with and without markups\. Figure[6](https://arxiv.org/html/2610.11520#A1.F6)shows a similar illustration by an example from the Naija treebank\. We hypothesize that recognition errors for segmentation symbols arise because these markups do not correspond to identifiable sounds in the audio signal\.

[Caron et al\. \(2019\)](https://arxiv.org/html/2610.11520#bib.bib1)found the punctuation marks useful for parsing, which can be because they used gold transcriptions, an ideal case when parsing transcriptions\.

We then assess if having punctuation in the training dataset helps parser to do a better demarcation \(comparing the last 2 setups\)\. The 2ndand 3rdrows in Table[6](https://arxiv.org/html/2610.11520#A1.T6)show this comparison\. The parsing performance remains same or improved with no punctuation marks in the training dataset, and thus we consider the Naija dataset without punctuation for all our evaluations\.

TreesAvg markupsin ground\_truthAvg markupsin prediction\[0, 2\]4231\.51\.966\.066\.00\.0\[3, 4\]2323\.43\.563\.064\.5\-1\.4\[5, 6\]1375\.44\.857\.161\.1\-3\.9\[7, 12\]1398\.77\.154\.260\.3\-6\.1\>=134116\.512\.644\.054\.0\-10\.0

Table 8:Parsing results comparison with and without syntax\-prosody segmentation annotations\.
## Appendix CSlovenian dataset

Learning from our findings on the Naija punctuation marks, we have removed the punctuations from the Slovenian dataset for our analysis\. Using the current version of the UD Slovenian dataset, we observed a low parsing performance \(30% LAS\)\. After investigation, we identified two types of noise, which we filtered out\. Additionally, we filtered out noisy recordings\.

1. 1\.Speech misalignments noise: Multiple transcripts are sometimes associated with a single speech recording without start or end timestamps\.
2. 2\.Transcription artefacts noise: The transcripts include placeholder tokens for personal names \(eg:\[name:personal\], \[name:surname\]\), which mislead the ASR system and consequently degrade parsing performance\. The corresponding audio hasbeepsfor it\.

For unique recordings, the speech–transcript alignment is not fully accurate, and we used forced alignment\([Pratap et al\., 2023](https://arxiv.org/html/2610.11520#bib.bib32)\)for it\.

## Appendix DAdditional Results

### D\.1Statistical Significance

Along with the parsing measures comparison, we perform significance testing on the test dataset of all 3 treebanks\. We use the non\-parametric Wilcoxon signed rank test\([Demšar, 2006](https://arxiv.org/html/2610.11520#bib.bib33)\)with the parsing measure LAS score as the measurement\. We first check for thenullhypothesis that there is no statistically significant difference between the performance of the simplified and the baseline architectures\. The 1strow in Table[10](https://arxiv.org/html/2610.11520#A5.T10)has the p\-value for this null hypothesis\. For Slovenian and ParisStories, we reject the null hypothesis at a confidence level of 5%, and conclude that there is a difference in the performance of 2 architectures\. For Naija, this null hypothesis cannot be rejected\.

For Slovenian and ParisStories, following that simplified and baseline architectures are statistically different, we checked for analternativehypothesis that the LAS score of ‘simplified’ architecture is greater than the ‘baseline’ architecture \(2ndrow in Table[10](https://arxiv.org/html/2610.11520#A5.T10)\)\. Here, the null hypothesis is that the difference in LAS score between simplified and baseline is less than a distribution symmetrical about zero\. For these 2 treebanks, this null hypothesis can be rejected at a confidence level of 5% , in favor of the alternative that the difference in LAS score between simplified and baseline architecture is greater than the distribution symmetric about zero\.

IndexHypothesisp valuep valuep valueNaijaSlovenianParisStories1Null21\.6%0\.2%3\.6%2Alternative\-0\.1%1\.8%

Table 9:Statistical significance testing between Simplified and Baseline architecture

## Appendix EComputational Time

All experiments were conducted on an NVIDIA H100 NVL GPU \(96GB memory\) with 64 cores\. The total computational budget for this paper \(including preliminary experiments\) was 800 GPU hours\.

The computation times for training and inference with the baseline and simplified Wav2tree on the Naija dataset are presented in Table[11](https://arxiv.org/html/2610.11520#A5.T11)\. The simplified architecture \(ours\) demonstrates a significant reduction in training time compared to the baseline architecture\([Pupier et al\., 2024](https://arxiv.org/html/2610.11520#bib.bib19), Wav2tree\)\. On the Naija corpus, and under identical hyperparameter settings, the simplified version requires only 60% of the training computational time needed by the baseline architecture\. The inference time is also reduced by 20% with the simplified architecture\.

Furthermore, with the simplified architecture, using intermediate layers of Wav2Vec2 \(21 layers\) results in an 5% reduction in training time compared to using the full set of layers\.

IndexHypothesisp valuep valuep valueNaijaSlovenianParisStories1Null21\.6%0\.2%3\.6%2Alternative\-0\.1%1\.8%

Table 10:Statistical significance testing between Simplified and Baseline architectureWav2treeWav2Vec2TrainingInferenceBaseline24 layers676min25sSimplified24 layers403min20sSimplified21 layers381min20s

Table 11:Computation time on the Naija corpus\.