GRAFT: Grafted Reference Audio for Fine-grained Pronunciation in Zero-shot Text-to-Speech

# Grafted Reference Audio for Fine-grained Pronunciation in Zero-shot Text-to-Speech
Source: [https://arxiv.org/html/2607.02633](https://arxiv.org/html/2607.02633)
Antonis Asonitis1,2,\*, Francesco Verdini1,3,\*, Aref Farhadipour1,4, Vijeta Avijeet1, Pierre\-Edouard Honnet1, Marzieh Razavi1, Juan Pablo Zuluaga Gomez1

###### Abstract

We present GRAFT, a per\-word pronunciation conditioning mechanism for text\-to\-speech neural codec language modeling\. Existing systems reach high intelligibility and naturalness but inherit the ambiguity of text and mispronounce rare proper nouns, loanwords and technical terms\. Even phoneme\-conditioned models offer no direct acoustic handle for per\-word pronunciation\. GRAFT controls the pronunciation of a chosen word from a short spoken sample of it, encoded with the model’s own speech tokenizer and bound to the word’s position in the prompt\. Voice conversion during training\-data construction disentangles the hint speaker from the target speaker, so the hint may come from any voice while the output stays in the target voice\. In a blind English listening study, human raters rank GRAFT first by a clear margin, judging its rendering of the difficult word closest to a reference recording of that word\. On a five\-language objective benchmark, GRAFT reduces target\-word phoneme error rate by 22–39% over the identical text\-only backbone and outperforms competitive open\-source zero\-shot systems, both phoneme\- and text\-conditioned, on target\-word pronunciation, while preserving speaker similarity and naturalness\.

## IIntroduction

Early end\-to\-end text\-to\-speech systems mapped a phoneme sequence to a waveform or a mel spectrogram\[[38](https://arxiv.org/html/2607.02633#bib.bib1),[35](https://arxiv.org/html/2607.02633#bib.bib2)\]\. The explicit phonetic intermediate made the learning problem well posed, since each phoneme corresponds to a small and largely consistent set of acoustic realisations\.

Recent systems extend a pretrained large language model with a vocabulary of discrete speech tokens produced by a neural audio codec, and continue training on paired text and speech\[[42](https://arxiv.org/html/2607.02633#bib.bib5),[47](https://arxiv.org/html/2607.02633#bib.bib10),[33](https://arxiv.org/html/2607.02633#bib.bib11),[10](https://arxiv.org/html/2607.02633#bib.bib9)\]\. The strong text prior inherited from the language model accelerates convergence and improves intelligibility on common words\. The downside is that text is a highly compressed representation of what is spoken\. The orthographic form fixes the identity of the word but underdetermines prosody, intonation, and the specific pronunciation in cases where more than one is valid\. For proper nouns, brand names, loanwords and technical terms, the text prior often fails to faithfully align with the target acoustics\.

![Refer to caption](https://arxiv.org/html/2607.02633v1/assets/teaser.png)Figure 1:GRAFT at inference\. A spoken example of the word \(“GRAFT”\), from any speaker, is silence\-trimmed \(VAD\), encoded to codec tokens, and placed at that word’s position in the text prompt\. Conditioned on a separate target\-speaker embedding, GRAFT synthesises the sentence in the target voice with the given pronunciation\.Models allowing jointly text and phonemes recover some control but do not fully resolve the problem\. Most pronunciation control relies on rule\-based grapheme\-to\-phoneme \(G2P\) lexicons\[[4](https://arxiv.org/html/2607.02633#bib.bib20),[12](https://arxiv.org/html/2607.02633#bib.bib21)\], which assume a single canonical pronunciation per word and cannot capture words that admit several valid realisations\. Such lexicons also generalise poorly to unseen words, the zero\-shot case that matters most for rare names and loanwords\. Neural phone recognisers such as ZIPA\[[49](https://arxiv.org/html/2607.02633#bib.bib41)\]and Wav2Vec2\-Phoneme\[[45](https://arxiv.org/html/2607.02633#bib.bib42)\]instead learn acoustic\-to\-phonetic mappings that generalise to new words and can convert an arbitrary recording into a phoneme sequence to prompt a phoneme\-conditioned TTS\. Even though this method offers tighter control, the fact that these phonemizers discard diacritical marks means the capacity to inject tonality or prosody is once again lost\.

A more direct option is to condition on the least compressed representation that the architecture of a text\-to\-speech model already accepts, namely the model’s own speech tokens\. Controlling a word’s pronunciation requires only a short recording of that word from any speaker\. This needs no phonetic notation or lexicon entry, so a non\-expert can prescribe a pronunciation simply by saying the word, and an acoustic example can also preserve the stress and tone that symbolic transcriptions discard\. The TTS should then render the full target phrase in the target speaker voice while preserving the given pronunciation of the chosen word\. We call this approach*GRAFT*\(Fig\.[1](https://arxiv.org/html/2607.02633#S1.F1)\)\.

Achieving this requires the model to disentangle the pronunciation in the reference recording from its speaker identity\. The reference clip carries both, but only the pronunciation should transfer, while the target voice comes from a separate reference\. This disentanglement follows from training on voice\-conversion pairs, whose objective forces the model to carry the content across speakers and to ignore the speaker identity of the source\. Once learned, the same disentanglement lets the model accept a per\-word hint from any speaker and render the target phrase in a different target voice\.

Our contributions are:

- •GRAFT, a per\-word pronunciation control for text\-to\-speech neural codec language modeling that requires no additional parameters or architectural change, where a spoken example of a word replaces its text tokens and guides its pronunciation in any target voice across five different languages\.111Samples available athttps://graft\-tts\.github\.io/graft/
- •Voice\-converted training data is the key enabling factor, disentangling the hint’s speaker from its pronunciation so a hint from any voice is rendered in the target voice\.
- •A multilingual difficult\-word benchmark \(five languages, openly licensed isolated references\), with released checkpoints, code and evaluation tooling\.

## IIRelated Work

### II\-ANeural codec language model TTS

The dominant paradigm casts speech synthesis as next\-token prediction over discrete codec tokens, established by VALL\-E\[[42](https://arxiv.org/html/2607.02633#bib.bib5)\]and scaled by many successors\[[39](https://arxiv.org/html/2607.02633#bib.bib6),[17](https://arxiv.org/html/2607.02633#bib.bib7),[22](https://arxiv.org/html/2607.02633#bib.bib8),[10](https://arxiv.org/html/2607.02633#bib.bib9),[44](https://arxiv.org/html/2607.02633#bib.bib16),[47](https://arxiv.org/html/2607.02633#bib.bib10),[33](https://arxiv.org/html/2607.02633#bib.bib11)\]\. Most reuse a pretrained text\-only LLM and extend its vocabulary with codec indices\[[47](https://arxiv.org/html/2607.02633#bib.bib10),[33](https://arxiv.org/html/2607.02633#bib.bib11)\], which is the source of both the strong text prior and the pronunciation failure modes that GRAFT targets\. They clone a voice zero\-shot by conditioning on a short acoustic example\[[18](https://arxiv.org/html/2607.02633#bib.bib44)\]\. GRAFT inherits this architecture unchanged and adds a per\-word audio conditioning slot consumed at inference, with no new parameters: the hint is encoded by the model’s existing codec tokenizer and spliced into the prompt, leaving inference cost unchanged apart from the few extra tokens\. Closest to our mechanism, WESCON\[[43](https://arxiv.org/html/2607.02633#bib.bib45)\]splices a reference acoustic prompt at a chosen word to control its emotional expression\. GRAFT applies similar per\-word conditioning for pronunciation and adds voice\-conversion training so the example may come from a different speaker than the target\.

### II\-BPhoneme\-conditioned TTS

Phoneme conditioning is the classical interface for pronunciation control, from sequence\-to\-sequence models\[[38](https://arxiv.org/html/2607.02633#bib.bib1),[35](https://arxiv.org/html/2607.02633#bib.bib2),[19](https://arxiv.org/html/2607.02633#bib.bib3),[20](https://arxiv.org/html/2607.02633#bib.bib4)\]to neural codec language models that mix phonemes into the prompt\[[17](https://arxiv.org/html/2607.02633#bib.bib7),[22](https://arxiv.org/html/2607.02633#bib.bib8)\]\. All commit to a fixed phonemic vocabulary\[[4](https://arxiv.org/html/2607.02633#bib.bib20),[12](https://arxiv.org/html/2607.02633#bib.bib21)\], which restricts coverage\. Improved text\-side front\-ends\[[15](https://arxiv.org/html/2607.02633#bib.bib46),[50](https://arxiv.org/html/2607.02633#bib.bib47)\]predict pronunciations more accurately but still emit symbols from a fixed inventory\. Per\-word fixes read an external lexicon entry\[[14](https://arxiv.org/html/2607.02633#bib.bib48)\]or edit the model’s weights for the target word\[[41](https://arxiv.org/html/2607.02633#bib.bib49)\], but both require a written phonetic specification\. GRAFT removes this commitment by conditioning on an instance of the desired pronunciation rather than a symbolic transcription\.

### II\-CForced alignment

Forced alignment locates word boundaries from a transcript, via HMM/GMM models such as the Montreal Forced Aligner\[[27](https://arxiv.org/html/2607.02633#bib.bib22)\]or CTC\-based neural aligners\[[21](https://arxiv.org/html/2607.02633#bib.bib23)\]\. A second family avoids a dedicated alignment model and instead recovers timing from the cross\-attention of an existing ASR or TTS decoder: the attention weights that link each output token to the audio frames are collapsed into a monotonic path with dynamic time warping, as in Whisper’s native word timestamps and tools such as whisper\-timestamped\[[26](https://arxiv.org/html/2607.02633#bib.bib34)\]and stable\-ts\[[16](https://arxiv.org/html/2607.02633#bib.bib35)\], a hybrid variant instead transfers timing from a separate CTC model onto the transcript\[[3](https://arxiv.org/html/2607.02633#bib.bib32)\]GRAFT uses the Qwen3\-TTS aligner\[[33](https://arxiv.org/html/2607.02633#bib.bib11)\]only in data preparation, to cut the per\-word reference clips\. None is needed at inference, where the input is already isolated\.

### II\-DVoice conversion

Voice conversion renders an utterance in a target speaker’s voice while preserving content, with recent zero\-shot systems including kNN\-VC\[[2](https://arxiv.org/html/2607.02633#bib.bib24)\], diffusion approaches\[[29](https://arxiv.org/html/2607.02633#bib.bib25)\]and Seed\-VC\[[24](https://arxiv.org/html/2607.02633#bib.bib26)\], which we use in data preparation\. Its speaker\-content factorisation has been studied via information\-bottleneck\[[31](https://arxiv.org/html/2607.02633#bib.bib50)\], self\-supervised content units\[[32](https://arxiv.org/html/2607.02633#bib.bib51)\]and perturbation training\[[8](https://arxiv.org/html/2607.02633#bib.bib52)\]\. Prior work treats voice conversion as the goal\. GRAFT uses it only for training data construction, in order to convert each hint to a different speaker, which teaches GRAFT to copy a word’s pronunciation and prosody while ignoring the hint speaker, so at inference the reference may come from any voice\. In this work, we show that this disentanglement is necessary to achieve natural\-sounding speech\.

## IIIMethod

### III\-APreliminaries

The base system is a neural codec language model that autoregressively predictsNN\-codebook codec tokens from a serialised text prompt\. We write a token sequence as𝐜∈ℤT×N\\mathbf\{c\}\\in\\mathbb\{Z\}^\{T\\times N\}\(TTframes atffHz\)\. We useQwen3\-TTS\-12Hz\-0\.6B\-Base\[[33](https://arxiv.org/html/2607.02633#bib.bib11)\]throughout \(N=16N\{=\}16,f=12\.5f\{=\}12\.5Hz\)\. It exposes a codec encoderEnc​\(⋅\)\\mathrm\{Enc\}\(\\cdot\)\(𝐜=Enc​\(𝐚\)\\mathbf\{c\}=\\mathrm\{Enc\}\(\\mathbf\{a\}\)on a 24 kHz waveform\), a decoderDec​\(⋅\)\\mathrm\{Dec\}\(\\cdot\), a text tokenizer, and special tokens\. GRAFT uses two slot delimiters,⟨Q⟩/⟨/Q⟩\\langle Q\\rangle/\\langle/Q\\rangle\(style references\) and⟨B⟩/⟨/B⟩\\langle B\\rangle/\\langle/B\\rangle\(per\-word grafts\), with a placeholder□\\square\.

TABLE I:Training mix by language\. Counts are pooled over the corpora feeding each language and use k \(10310^\{3\}\) and M \(10610^\{6\}\)\. Corpora: VCTK\[[46](https://arxiv.org/html/2607.02633#bib.bib27)\]\(V\), LibriTTS\[[48](https://arxiv.org/html/2607.02633#bib.bib28)\]\(L\), MLS\[[30](https://arxiv.org/html/2607.02633#bib.bib29)\]\(M\) and Common Voice\[[1](https://arxiv.org/html/2607.02633#bib.bib30)\]\(C\)\. “\#Spk” is unique speaker IDs, “\#Utt” utterances after preprocessing\.Lang\.Corpora\#SpkHours\#UttENV, L, C12\.912\.9k1\.51\.5k1\.11\.1MDEM, C0\.80\.8k2\.02\.0k0\.90\.9MFRM, C9\.09\.0k1\.91\.9k0\.80\.8MESM, C4\.94\.9k1\.41\.4k0\.60\.6MITM, C1\.51\.5k0\.50\.5k0\.20\.2MTotal—28\.4\\mathbf\{28\.4\}k7\.3\\mathbf\{7\.3\}k3\.2\\mathbf\{3\.2\}M
### III\-BPer\-word audio binding

GRAFT conditions on a tuple\(t,𝐫,\{\(wk,𝐚k\)\}k=1K\)\\bigl\(t,\\,\\mathbf\{r\},\\,\\\{\(w\_\{k\},\\mathbf\{a\}\_\{k\}\)\\\}\_\{k=1\}^\{K\}\\bigr\)wherettis the target text,𝐫\\mathbf\{r\}is a target\-voice reference clip and each\(wk,𝐚k\)\(w\_\{k\},\\mathbf\{a\}\_\{k\}\)is a target word paired with an isolated audio recording of that word\. The reference clip and the per\-word clips are encoded by the base model’s own codec,𝐜ref=Enc​\(𝐫\)\\mathbf\{c\}^\{\\mathrm\{ref\}\}=\\mathrm\{Enc\}\(\\mathbf\{r\}\)of lengthTrT\_\{r\}, and𝐜\(k\)=Enc​\(𝐚k\)\\mathbf\{c\}^\{\(k\)\}=\\mathrm\{Enc\}\(\\mathbf\{a\}\_\{k\}\)of lengthTkT\_\{k\}\.

Let the target text bet=m1​⋯​mMt=m\_\{1\}\\cdots m\_\{M\}and𝒢=\{i:mi=wk​for some​k\}\\mathcal\{G\}=\\\{i:m\_\{i\}=w\_\{k\}\\text\{ for some \}k\\\}index the grafted words\. The prompt concatenates a fixed role prefix, the style slot\[⟨Q⟩,□Tr,⟨/Q⟩\]\[\\langle Q\\rangle,\\square^\{T\_\{r\}\},\\langle/Q\\rangle\], and a body in which each ungrafted word \(i∉𝒢i\\notin\\mathcal\{G\}\) keeps its text tokens while each grafted word \(i∈𝒢i\\in\\mathcal\{G\}\) contributes the graft slot\[⟨B⟩,□Tk,⟨/B⟩\]\[\\langle B\\rangle,\\square^\{T\_\{k\}\},\\langle/B\\rangle\]in place\. A grafted word is thus represented by its codec tokens*instead of*its text, never in addition to it\. The placeholder text embeddings are replaced at the corresponding positions by the codec embeddings𝐞tcodec=∑n=0N−1En​\(𝐜​\[t,n\]\)\\mathbf\{e\}^\{\\mathrm\{codec\}\}\_\{t\}=\\sum\_\{n=0\}^\{N\-1\}E\_\{n\}\\\!\\bigl\(\\mathbf\{c\}\[t,n\]\\bigr\), summed across allN=16N\{=\}16residual codebooks, whereEnE\_\{n\}is the embedding table for thenn\-th codebook\. The talker transformer then generates the remaining codec frames autoregressively, conditioned on this composite input\.

### III\-CTraining data construction

![Refer to caption](https://arxiv.org/html/2607.02633v1/assets/training.png)Figure 2:Training\-data construction\. From a Speaker A utterance we make a voice\-converted copy in a different voice, force\-align it to find word boundaries, and re\-encode the target word \(“Nguyen”\) with the codec as a per\-word graft\. The graft replaces that word’s text in the prompt, and the model is trained to predict the codec of the*original*Speaker A utterance\. Since the graft is in a different voice from the target, the model learns to copy pronunciation while ignoring the hint speaker\.We construct training samples that match the inference\-time use of GRAFT, using only the public read\-speech and crowd\-sourced corpora of Table[I](https://arxiv.org/html/2607.02633#S3.T1)and no additional human\-supplied audio\.

For every training utteranceuuwith original waveform𝐚u\\mathbf\{a\}^\{u\}and transcripttut^\{u\}, we produce a parallel voice\-converted version𝐚^u=SeedVCD​\(𝐚u,𝐫u\)\\hat\{\\mathbf\{a\}\}^\{u\}=\\mathrm\{SeedVC\}\_\{D\}\(\\mathbf\{a\}^\{u\},\\mathbf\{r\}^\{u\}\), where𝐫u\\mathbf\{r\}^\{u\}is a reference clip from a random other speaker and Seed\-VC\[[24](https://arxiv.org/html/2607.02633#bib.bib26)\]runs withD=20D=20steps, so the utterance is said in a different voice\. We force\-align𝐚^u\\hat\{\\mathbf\{a\}\}^\{u\}totut^\{u\}, cut each wordii, pad it with0\.30\.3s of silence on each side, encode it with the base codec, and discard the codec frames in the padding\. The result is a per\-word codec sequence𝐜u,i∈ℤTu,i×N\\mathbf\{c\}^\{u,i\}\\\!\\in\\\!\\mathbb\{Z\}^\{T^\{u,i\}\\times N\}for every word, each in a speaker different from the original\.

The training target is always the original waveform’s codec,𝐜u,tgt=Enc​\(𝐚u\)\\mathbf\{c\}^\{u,\\mathrm\{tgt\}\}=\\mathrm\{Enc\}\(\\mathbf\{a\}^\{u\}\)\. The voice\-converted utterance is used as input for per\-word reference clips spliced into the prompt and never as target\. Therefore, GRAFT’s output quality is not limited by the voice conversion model\.

TABLE II:Ablation of training\-data choices, evaluated on a separate set of100100difficult English words manually selected from the ACL 60/60 corpus\[[37](https://arxiv.org/html/2607.02633#bib.bib31)\]\.*no VC*removes voice conversion from training\-data construction and*no aug\.*removes the per\-word hint augmentations\. SSIM is speaker similarity to the target voice\. Best per column in bold\.∗no\-VC reaches the lowest D\-PER only by copying the hint tokens straight through, which collapses speaker similarity, so it is a degenerate solution rather than the strongest system\.SystemD\-PER↓\\downarrowUTMOS↑\\uparrowWER↓\\downarrowSSIM↑\\uparrowGRAFT \(no aug\.\)0\.250\.254\.214\.210\.080\.080\.670\.67GRAFT \(no VC\)0\.16∗0\.16^\{\*\}3\.893\.890\.180\.180\.200\.20GRAFT0\.190\.194\.43\\mathbf\{4\.43\}0\.05\\mathbf\{0\.05\}0\.68\\mathbf\{0\.68\}TABLE III:Difficult\-word benchmark \(500500words per language, one run per phrase\)\. Phoneme baselines \(StyleTTS2, MaskGCT, Matcha\-TTS, Piper\) receive the hint*audio*transcribed to espeak phonemes \(wav2vec2\-espeak\[[45](https://arxiv.org/html/2607.02633#bib.bib42)\]\), the same hint pronunciation GRAFT gets but in symbolic form\. Text\-input systems \(XTTS\-v2, F5\-TTS, CosyVoice2\) get only the spelled carrier, and GRAFT gets the hint as audio\. Systems appear only for languages they support\. D\-PER is mean±\{\\pm\}std over words, and D\-PER10%averages the hardest10%10\\%\.↓\\downarrow/↑\\uparrowlower/higher is better\.†single\-speaker \(excluded from the SSIM ranking\)\. Bestbold, secondunderlined\. The English Bradley\-Terry study is in Fig\.[IV](https://arxiv.org/html/2607.02633#S5.T4)\.PronunciationGeneralSystemD\-PER↓\\downarrowD\-PER10%↓\\downarrowWS↑\\uparrowWER↓\\downarrowUTMOS↑\\uparrowSSIM↑\\uparrowEnglishQwen3\-TTS\-0\.6B0\.37¯±0\.42\\underline\{0\.37\}\\,\{\\scriptstyle\\pm 0\.42\}1\.161\.160\.91¯\\underline\{0\.91\}0\.02\\mathbf\{0\.02\}4\.46¯\\underline\{4\.46\}0\.730\.73StyleTTS2 \(ph\.\)0\.65±0\.280\.65\\,\{\\scriptstyle\\pm 0\.28\}1\.111\.110\.830\.830\.170\.174\.354\.350\.620\.62MaskGCT \(ph\.\)0\.46±0\.340\.46\\,\{\\scriptstyle\\pm 0\.34\}1\.151\.150\.880\.880\.170\.174\.224\.220\.80\\mathbf\{0\.80\}Matcha \(ph\.\)0\.43±0\.340\.43\\,\{\\scriptstyle\\pm 0\.34\}1\.111\.110\.890\.890\.060\.064\.344\.340\.15†0\.15^\{\\dagger\}Piper \(ph\.\)0\.48±0\.360\.48\\,\{\\scriptstyle\\pm 0\.36\}1\.121\.120\.860\.860\.070\.074\.344\.340\.12†0\.12^\{\\dagger\}XTTS\-v20\.41±0\.380\.41\\,\{\\scriptstyle\\pm 0\.38\}1\.201\.200\.900\.900\.02\\mathbf\{0\.02\}4\.194\.190\.670\.67F5\-TTS0\.38±0\.400\.38\\,\{\\scriptstyle\\pm 0\.40\}1\.10¯\\underline\{1\.10\}0\.890\.890\.02\\mathbf\{0\.02\}4\.354\.350\.79¯\\underline\{0\.79\}CosyVoice20\.39±0\.390\.39\\,\{\\scriptstyle\\pm 0\.39\}1\.161\.160\.900\.900\.03¯\\underline\{0\.03\}4\.46¯\\underline\{4\.46\}0\.710\.71GRAFT0\.29±0\.16\\mathbf\{0\.29\}\\,\{\\scriptstyle\\pm 0\.16\}0\.81\\mathbf\{0\.81\}0\.92\\mathbf\{0\.92\}0\.090\.094\.47\\mathbf\{4\.47\}0\.760\.76GermanQwen3\-TTS\-0\.6B0\.49¯±0\.33\\underline\{0\.49\}\\,\{\\scriptstyle\\pm 0\.33\}1\.13¯\\underline\{1\.13\}0\.90¯\\underline\{0\.90\}0\.07\\mathbf\{0\.07\}4\.03\\mathbf\{4\.03\}0\.69¯\\underline\{0\.69\}MaskGCT \(ph\.\)0\.56±0\.380\.56\\,\{\\scriptstyle\\pm 0\.38\}1\.281\.280\.870\.870\.170\.173\.073\.070\.79\\mathbf\{0\.79\}Piper \(ph\.\)0\.55±0\.330\.55\\,\{\\scriptstyle\\pm 0\.33\}1\.191\.190\.870\.870\.130\.133\.363\.360\.14†0\.14^\{\\dagger\}XTTS\-v20\.50±0\.390\.50\\,\{\\scriptstyle\\pm 0\.39\}1\.251\.250\.90¯\\underline\{0\.90\}0\.11¯\\underline\{0\.11\}3\.183\.180\.69¯\\underline\{0\.69\}GRAFT0\.30±0\.23\\mathbf\{0\.30\}\\,\{\\scriptstyle\\pm 0\.23\}0\.76\\mathbf\{0\.76\}0\.94\\mathbf\{0\.94\}0\.11¯\\underline\{0\.11\}3\.99¯\\underline\{3\.99\}0\.680\.68FrenchQwen3\-TTS\-0\.6B0\.41¯±0\.37\\underline\{0\.41\}\\,\{\\scriptstyle\\pm 0\.37\}1\.201\.200\.89¯\\underline\{0\.89\}0\.05\\mathbf\{0\.05\}3\.64¯\\underline\{3\.64\}0\.700\.70MaskGCT \(ph\.\)0\.56±0\.350\.56\\,\{\\scriptstyle\\pm 0\.35\}1\.251\.250\.860\.860\.180\.182\.802\.800\.81\\mathbf\{0\.81\}Piper \(ph\.\)0\.55±0\.340\.55\\,\{\\scriptstyle\\pm 0\.34\}1\.231\.230\.850\.850\.140\.143\.403\.400\.11†0\.11^\{\\dagger\}XTTS\-v20\.43±0\.340\.43\\,\{\\scriptstyle\\pm 0\.34\}1\.11¯\\underline\{1\.11\}0\.880\.880\.06¯\\underline\{0\.06\}3\.043\.040\.700\.70GRAFT0\.28±0\.18\\mathbf\{0\.28\}\\,\{\\scriptstyle\\pm 0\.18\}0\.77\\mathbf\{0\.77\}0\.92\\mathbf\{0\.92\}0\.070\.073\.72\\mathbf\{3\.72\}0\.71¯\\underline\{0\.71\}SpanishQwen3\-TTS\-0\.6B0\.38¯±0\.32\\underline\{0\.38\}\\,\{\\scriptstyle\\pm 0\.32\}1\.021\.020\.90¯\\underline\{0\.90\}0\.02\\mathbf\{0\.02\}3\.76¯\\underline\{3\.76\}0\.69\\mathbf\{0\.69\}Piper \(ph\.\)0\.56±0\.330\.56\\,\{\\scriptstyle\\pm 0\.33\}1\.161\.160\.850\.850\.100\.102\.532\.530\.08†0\.08^\{\\dagger\}XTTS\-v20\.40±0\.290\.40\\,\{\\scriptstyle\\pm 0\.29\}0\.99¯\\underline\{0\.99\}0\.890\.890\.04¯\\underline\{0\.04\}3\.083\.080\.650\.65GRAFT0\.29±0\.20\\mathbf\{0\.29\}\\,\{\\scriptstyle\\pm 0\.20\}0\.68\\mathbf\{0\.68\}0\.91\\mathbf\{0\.91\}0\.090\.093\.80\\mathbf\{3\.80\}0\.66¯\\underline\{0\.66\}ItalianQwen3\-TTS\-0\.6B0\.40±0\.280\.40\\,\{\\scriptstyle\\pm 0\.28\}0\.970\.970\.89¯\\underline\{0\.89\}0\.06\\mathbf\{0\.06\}3\.68¯\\underline\{3\.68\}0\.70¯\\underline\{0\.70\}Piper \(ph\.\)0\.58±0\.270\.58\\,\{\\scriptstyle\\pm 0\.27\}1\.131\.130\.810\.810\.150\.153\.403\.400\.13†0\.13^\{\\dagger\}XTTS\-v20\.38¯±0\.27\\underline\{0\.38\}\\,\{\\scriptstyle\\pm 0\.27\}0\.94¯\\underline\{0\.94\}0\.880\.880\.09¯\\underline\{0\.09\}2\.972\.970\.71\\mathbf\{0\.71\}GRAFT0\.25±0\.17\\mathbf\{0\.25\}\\,\{\\scriptstyle\\pm 0\.17\}0\.66\\mathbf\{0\.66\}0\.91\\mathbf\{0\.91\}0\.110\.113\.73\\mathbf\{3\.73\}0\.71\\mathbf\{0\.71\}To assemble a training input, we walk over the words oftut^\{u\}and decide for each word whether to insert a per\-word graft\. A word is eligible if its aligned duration is at least0\.040\.04s and its surface form is at least four characters long\. Each eligible wordiiis selected independently with probabilityphint=0\.5p\_\{\\mathrm\{hint\}\}=0\.5\. For every selected word, the actual graft codec sequence is sampled according to

𝐜\(i\)=\{𝐜u,iw\.p\.​\(1−ps\)​\(1−pa\),Aug​\(𝐜u,i\)w\.p\.​\(1−ps\)​pa,𝐜u′,iw\.p\.​ps​\(1−pa\),Aug​\(𝐜u′,i\)w\.p\.​ps​pa,\\mathbf\{c\}^\{\(i\)\}\\;=\\;\\begin\{cases\}\\mathbf\{c\}^\{u,i\}&\\text\{w\.p\. \}\(1\{\-\}p\_\{s\}\)\(1\{\-\}p\_\{a\}\),\\\\ \\mathrm\{Aug\}\\bigl\(\\mathbf\{c\}^\{u,i\}\\bigr\)&\\text\{w\.p\. \}\(1\{\-\}p\_\{s\}\)\\,p\_\{a\},\\\\ \\mathbf\{c\}^\{u^\{\\prime\},i\}&\\text\{w\.p\. \}p\_\{s\}\(1\{\-\}p\_\{a\}\),\\\\ \\mathrm\{Aug\}\\bigl\(\\mathbf\{c\}^\{u^\{\\prime\},i\}\\bigr\)&\\text\{w\.p\. \}p\_\{s\}\\,p\_\{a\},\\end\{cases\}\(1\)withps=0\.3p\_\{s\}=0\.3andpa=0\.5p\_\{a\}=0\.5\. The first two cases keep the voice\-converted clip from the current utterance, with or without augmentation\. The last two swap in the same word from a different utteranceu′u^\{\\prime\}\. The swap matters because an in\-context graft is cut from the very sentence being predicted and is therefore coarticulated and prosodically consistent with the carrier, whereas at inference the reference is an isolated recording\. Sourcing the graft from a different utterance breaks this coarticulation and matches the isolated\-word prompts seen at inference\.

Augmented variants are precomputed offline: each per\-word clip is decoded, perturbed, and re\-encoded, soAug​\(𝐜\)=Enc​\(perturb​\(Dec​\(𝐜\)\)\)\\mathrm\{Aug\}\(\\mathbf\{c\}\)=\\mathrm\{Enc\}\(\\mathrm\{perturb\}\(\\mathrm\{Dec\}\(\\mathbf\{c\}\)\)\)in Eq\. \([1](https://arxiv.org/html/2607.02633#S3.E1)\)\. The perturbations cover additive noise, level and loudness jitter, pitch and speed changes, telephone\-band filtering, compression, reverberation and variable lead/trail silence, alone and in combinations emulating phone, quiet\-room, far\-microphone and loud\-close conditions\. They broaden the conditioning distribution along the two axes that matter at inference: imperfect word boundaries, where the variable lead/trail silence makes the model robust to spans that are cut loosely, and consumer\-microphone recording in noisy rooms\.

### III\-DTraining objective

We fine\-tune the base model with the standard next\-token cross\-entropy loss on the talker codec tokens plus an auxiliary sub\-talker loss weightedλsub=0\.3\\lambda\_\{\\mathrm\{sub\}\}=0\.3,ℒ=ℒtalker\+λsub​ℒsub−talker\\mathcal\{L\}=\\mathcal\{L\}\_\{\\mathrm\{talker\}\}\+\\lambda\_\{\\mathrm\{sub\}\}\\,\\mathcal\{L\}\_\{\\mathrm\{sub\-talker\}\}\. Training runs over the full mix of Table[I](https://arxiv.org/html/2607.02633#S3.T1)with effective batch size192192, AdamW\[[25](https://arxiv.org/html/2607.02633#bib.bib36)\], learning rate5×10−65\{\\times\}10^\{\-6\}, weight decay0\.010\.01, a cosine schedule and0\.020\.02warmup, taking a total of 48 hours on a single 96 GB GPU\. The ablations use the identical schedule and data count, isolating the training\-data choice\.

### III\-EInference

At inference the inputs are onlytt,𝐫\\mathbf\{r\}and the\(wk,𝐚k\)\(w\_\{k\},\\mathbf\{a\}\_\{k\}\)pairs, with no forced alignment since each𝐚k\\mathbf\{a\}\_\{k\}is already isolated\. We use standard autoregressive sampling, and trimming each clip with a voice activity detector \(VAD\)\[[40](https://arxiv.org/html/2607.02633#bib.bib43)\]before encoding brings it closer to the training distribution\.

## IVExperimental Setup

### IV\-ADifficult\-word benchmark

We evaluate on a new multilingual benchmark of difficult words\. For each of five languages \(English, German, French, Spanish, Italian\) we select the rarest words bywordfreqZipf score \(band1\.31\.3–3\.33\.3\) from the openly licensed Wikimedia Lingua Libre collection of isolated single\-word recordings \(500500per language,2,5002\{,\}500total\), place each in a natural carrier phrase, and use its clean isolated human recording as the per\-word hint\. Because the recordings are isolated, no forced alignment or cropping is needed, and every item is openly licensed \(CC0/CC\-BY/CC\-BY\-SA\), so the benchmark is redistributed with its audio\. Each phrase is synthesised once per system\. We report target\-word D\-PER \(mean±\{\\pm\}std\) with its mean over the hardest10%10\\%, alongside WER, UTMOS and speaker similarity\.

### IV\-BCompared systems

The text\-only base, Qwen3\-TTS\-12Hz\-0\.6B\-Base\[[33](https://arxiv.org/html/2607.02633#bib.bib11)\], is the principal baseline, receiving only the target text and the style reference with no per\-word audio interface\. We additionally compare against seven open\-source zero\-shot systems, grouped by conditioning interface\. Four take a phonemic front\-end: StyleTTS2\[[23](https://arxiv.org/html/2607.02633#bib.bib14)\]\(diffusion\), MaskGCT\[[44](https://arxiv.org/html/2607.02633#bib.bib16)\]\(masked codec LM\), Matcha\-TTS\[[28](https://arxiv.org/html/2607.02633#bib.bib17)\]\(flow matching\) and Piper\[[13](https://arxiv.org/html/2607.02633#bib.bib18)\]\(VITS\)\. Three are text\-input: XTTS v2\[[5](https://arxiv.org/html/2607.02633#bib.bib15)\], F5\-TTS\[[7](https://arxiv.org/html/2607.02633#bib.bib13)\]and CosyVoice2\[[11](https://arxiv.org/html/2607.02633#bib.bib19)\]\. All receive the same carrier and speaker reference as GRAFT, so the only difference is the model and its interface\. The phoneme\-input systems do not merely get the spelled word: we transcribe the hint*audio*to espeak phonemes \(wav2vec2\-espeak\[[45](https://arxiv.org/html/2607.02633#bib.bib42)\]\) and substitute that for the target word, the matched symbolic counterpart to GRAFT’s audio hint\. The text\-input systems cannot consume phonemes and receive only the spelled carrier\. The methods designed specifically for per\-word correction, Neural Lexicon Reader\[[14](https://arxiv.org/html/2607.02633#bib.bib48)\]and SonoEdit\[[41](https://arxiv.org/html/2607.02633#bib.bib49)\], both condition on a*written*phonetic specification of the target word\. For our rare words this specification is absent, and where it exists their input reduces to the oracle\-phoneme condition \(Table[V](https://arxiv.org/html/2607.02633#S5.T5)\), which does not close the gap to GRAFT’s audio interface\.

To isolate the contribution of the two key training\-data choices, we compare GRAFT against two ablations of itself\.*GRAFT \(no VC\)*slices the per\-word hint clips directly from the source\-speaker recordings instead of voice\-converted copies, with the cross\-utterance swap and augmentation disabled so no converted content re\-enters\.*GRAFT \(no aug\.\)*encodes the hints clean, without the silence padding \(Section[III\-C](https://arxiv.org/html/2607.02633#S3.SS3)\) or the perturbationsAug​\(⋅\)\\mathrm\{Aug\}\(\\cdot\)of Eq\. \([1](https://arxiv.org/html/2607.02633#S3.E1)\)\.

### IV\-CMetrics

We use five objective metrics and one subjective metric\. WER is the sentence\-level word error rate between the carrier text and the Whisper\-large\-v3\[[34](https://arxiv.org/html/2607.02633#bib.bib37)\]transcript of the generated utterance, capped at1\.01\.0per item\. D\-PER, the main pronunciation metric, is the phoneme error rate between the hint clip and the targeted word cut from the generated audio: both are phonemised with the ZIPA universal phone recogniser\[[49](https://arxiv.org/html/2607.02633#bib.bib41)\]and the Levenshtein distance between the phone sequences is normalised by the reference length\. Naturalness is UTMOS\[[36](https://arxiv.org/html/2607.02633#bib.bib39)\]\. Whisper similarity \(WS\) is the cosine similarity between the Whisper\-encoder embeddings of the targeted word in the generated and hint clips, a content representation that largely discards speaker identity, so WS reflects pronunciation rather than voice\. Speaker preservation \(SSIM\) is the WavLM\[[6](https://arxiv.org/html/2607.02633#bib.bib38)\]cosine similarity to the cloned reference\. Finally, a blind pairwise listening study, run with an online tool,222https://www\.mabyduck\.com/gives per\-system Bradley\-Terry scores for English \(Fig\.[IV](https://arxiv.org/html/2607.02633#S5.T4)\)\.

## VResults

### V\-ADifficult\-word evaluation

Table[III](https://arxiv.org/html/2607.02633#S3.T3)reports all metrics on the benchmark of Section[IV](https://arxiv.org/html/2607.02633#S4), discussing English first and the other four languages in Section[V\-B](https://arxiv.org/html/2607.02633#S5.SS2)\. Because GRAFT shares the base model’s backbone and differs in the conditioning interface, the GRAFT\-versus\-base comparison largely isolates the effect of that interface\.

GRAFTXTTS\-v2F5\-TTSCosyVoice2Qwen3\-TTSMatchaPiperMaskGCTStyleTTS21,6001\{,\}6001,8001\{,\}8002,0002\{,\}0002,2002\{,\}200Bradley\-Terry score \(95%95\\%CI\)

Figure 3:Blind pairwise human listening study on English \(ratings initialised at10001000\)\. Thirty difficult words from the benchmark, judged by2020paid listeners over600600comparisons\. With a clean recording of the target word as reference, each listener was asked:*“which of the two sentences has this pronunciation closer in terms of phonetics, linguistics and prosody, while remaining natural?”*Bars are Bradley\-Terry scores, whiskers95%95\\%confidence intervals\. GRAFT \(green\) ranks first, leading the next system by150150points, with a lower bound \(21072107\) above every baseline point estimate\.TABLE IV:Several per\-word hints in one carrier, over100100multi\-target carriers\.nnis the grafted\-word count, D\-PER averaged over them\.HintsnnD\-PER↓\\downarrowWER↓\\downarrowUTMOS↑\\uparrow110\.2820\.2820\.0490\.0494\.4704\.470220\.3050\.3050\.1170\.1174\.4454\.445330\.3280\.3280\.2220\.2224\.4404\.440440\.3310\.3310\.2450\.2454\.4584\.458550\.3760\.3760\.2060\.2064\.4944\.494
On phonetic faithfulness GRAFT separates clearly from the baselines, and its D\-PER reduction over the base is significant in every language \(paired Wilcoxon over the500500words per language,p<10−7p<10^\{\-7\}\)\. It also roughly halves the per\-word D\-PER spread \(±0\.16\\pm 0\.16vs\.±0\.42\\pm 0\.42in English\), so the correction is more consistent, not only lower\. The Lingua Libre hints are reference single\-word recordings by fluent speakers, taken as ground\-truth\. Scoring against the hint invites two objections\. The acoustic one, that GRAFT pastes the hint through, is the no\-VC ablation: it reaches the*lowest*D\-PER \(0\.160\.16, Table[II](https://arxiv.org/html/2607.02633#S3.T2)\) only by copying the hint while speaker similarity collapses to0\.200\.20, a mode we train against\. The metric one, that D\-PER phonemises both sides with one recogniser \(ZIPA\) and GRAFT alone gets the hint as audio, so the score might reward channel match, is ruled out by two judges sharing neither GRAFT’s modality nor ZIPA: GRAFT also leads on WS \(a Whisper\-encoder distance\) and in the human study \(Fig\.[IV](https://arxiv.org/html/2607.02633#S5.T4), no phone metric\)\. The benchmark thus measures faithful transfer of the supplied pronunciation, the deployment case, not canonical correctness\. Giving the phoneme baselines the hint as phonemes narrows but does not close the gap, and a human\-curated dictionary pronunciation does not either \(Table[V](https://arxiv.org/html/2607.02633#S5.T5), English dictionary\-covered words\), ruling out transcription error as the cause rather than settling the rarer\-word case\. Because WER is sentence\-level, one mispronounced target word barely moves it, so a system can post a low WER while still getting the word wrong\. GRAFT’s higher WER is itself concentrated on the target word, removing it drops the English WER from0\.090\.09to0\.030\.03\(base0\.020\.02to0\.010\.01\) and the five\-language average from0\.090\.09to0\.050\.05, leaving the carrier over95%95\\%correct\. The residual is partly the recogniser’s, even on a clean isolated human recording, transcribed in isolation, Whisper\-large\-v3\[[34](https://arxiv.org/html/2607.02633#bib.bib37)\]returns the exact word only75%75\\%of the time in English \(66%66\\%overall\), and on over half of GRAFT’s misses the rendered pronunciation still matches the hint \(D\-PER≤0\.34\{\\leq\}0\.34\)\. Naturalness is largely unaffected: GRAFT leads the base on UTMOS in four of five languages\.

The ablations \(Table[II](https://arxiv.org/html/2607.02633#S3.T2)\) isolate voice\-conversion training\. To test whether the grafted word merely echoes the hint speaker, the generated target word is cut from each render with the same forced aligner, resampled to1616kHz, and scored with ECAPA\-TDNN\[[9](https://arxiv.org/html/2607.02633#bib.bib40)\]against both the hint speaker and the cloned target speaker, averaged over the benchmark items\. For the no\-VC ablation this confirms the paste\-through as the grafted word sits on the hint speaker \(0\.670\.67\) far more than on the target \(0\.100\.10\)\. With voice conversion the grafted word instead carries the target identity rather than the hint’s \(0\.470\.47versus0\.120\.12\), which a copy of the hint waveform cannot produce and which separates GRAFT’s low D\-PER from acoustic leakage\.

### V\-BMultilingual evaluation

The hint conditions on acoustic tokens rather than text or phonemes, so the interface carries no language\-specific machinery\. The non\-English rows of Table[III](https://arxiv.org/html/2607.02633#S3.T3)repeat the English pattern as GRAFT cuts target\-word D\-PER by2222–39%39\\%in every language and raises hint similarity, with naturalness and speaker similarity preserved at a small WER cost\.

TABLE V:Oracle\-pronunciation control \(English CMUdict\-covered words,n=225n=225\)\. Phoneme baselines receive either the wav2vec2\-espeak transcription of the hint or the human\-curated CMUdict pronunciation\. Neither closes the gap to GRAFT, which rules out transcription error as the cause of the baseline deficit on these covered words\. D\-PER is scored against the hint, which the audio\-derived phonemes track directly\.Systemhint phonemes↓\\downarroworacle \(CMUdict\)↓\\downarrowStyleTTS20\.650\.650\.690\.69MaskGCT0\.490\.490\.450\.45Matcha\-TTS0\.430\.430\.490\.49Piper0\.490\.490\.510\.51GRAFT\(audio\)0\.30\\mathbf\{0\.30\}

## VIDiscussion

The human study is GRAFT’s strongest evidence, a human judgment rather than a phone\-distance score\. The reference shown to raters is the same recording GRAFT receives as its hint, so the study measures which model reproduces a supplied pronunciation most faithfully, not which is canonically correct\. Every system is given that same pronunciation, as audio for GRAFT and as transcribed phonemes or spelling for the others\. Across3030difficult words and600600comparisons GRAFT wins outright, leading the strongest baseline by150150Bradley\-Terry points with its confidence\-interval lower bound above every other system’s point estimate\. It does so not by echoing the clip but by re\-rendering the pronunciation in the cloned target voice \(the per\-word identity above\), so the preference reflects pronunciation transfer, not playback of the hint\.

Because the hint is an example rather than a transcription, a non\-expert can correct a chosen word by saying it, and disentanglement keeps the output in the chosen voice\. This helps where text and phoneme front\-ends fail: rare proper nouns, loanwords and technical terms whose realisation is known to a speaker yet absent from any lexicon\.

Training grafts every eligible word \(Section[III\-C](https://arxiv.org/html/2607.02633#S3.SS3)\), so several hints stack in one carrier: up to five \(Table[IV](https://arxiv.org/html/2607.02633#S5.T4)\) raise per\-word D\-PER only gently \(0\.280\.28to0\.380\.38\) while carrier WER absorbs the seam cost \(0\.050\.05to over0\.200\.20\) and naturalness is unchanged, so the price of extra grafts falls on fluency, not pronunciation\.

The method is reliable when the hint matches the trimmed, neutral training distribution, and mistrimmed or low\-quality hints can yield a wrong pronunciation\. The interface operates on single words, and GRAFT is shown on one0\.60\.6B base model\. The benchmark is non\-tonal, so the measured gains reflect segmental and stress detail rather than the tonal advantage motivated against symbolic phonemizers, which discard diacritics\. Multi\-word spans, prosody and tonal\-language transfer, where such phonemizers should fail most, and scaling are left to future work\.

## VIIConclusion

GRAFT adds a per\-word audio conditioning slot to neural codec language model TTS: a short spoken example fixes a word’s pronunciation, and voice\-conversion training renders it in the cloned voice from any speaker\. The mechanism reuses the base model’s own tokenizer and adds no new parameters, so per\-word control drops into an existing system at the cost of only a few extra tokens at inference\. Because the interface is a spoken example rather than a symbolic transcription, a non\-expert can prescribe a pronunciation simply by saying the word, which is precisely what rare proper nouns, loanwords and technical terms demand and what text and phoneme front\-ends cannot supply\.

Human raters rank GRAFT first in a blind English listening study, and on the objective benchmark it improves target\-word pronunciation fidelity over the text\-only base \(D\-PER reduced by2222–39%39\\%\) with consistent gains across five languages, while preserving speaker similarity and naturalness\. We release the checkpoints, code and the multilingual difficult\-word benchmark to support further work\. Tonal\-language transfer, multi\-word spans, and scaling to larger backbones, where symbolic front\-ends should struggle most, are natural next steps\.

## References

- \[1\]R\. Ardila, M\. Branson, K\. Davis, M\. Kohler, J\. Meyer, M\. Henretty, R\. Morais, L\. Saunders, F\. Tyers, and G\. Weber\(2020\)Common voice: a massively\-multilingual speech corpus\.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\. 4218–4222\(English\)\.External Links:ISBN 979\-10\-95546\-34\-4,[Link](https://aclanthology.org/2020.lrec-1.520)Cited by:[TABLE I](https://arxiv.org/html/2607.02633#S3.T1)\.
- \[2\]M\. Baas, B\. van Niekerk, and H\. Kamper\(2023\)Voice conversion with just nearest neighbors\.In24th Annual Conference of the International Speech Communication Association, Interspeech 2023, Dublin, Ireland, August 20\-24, 2023,N\. Harte, J\. Carson\-Berndsen, and G\. Jones \(Eds\.\),pp\. 2053–2057\.External Links:[Document](https://dx.doi.org/10.21437/INTERSPEECH.2023-419),[Link](https://doi.org/10.21437/Interspeech.2023-419)Cited by:[§II\-D](https://arxiv.org/html/2607.02633#S2.SS4.p1.1)\.
- \[3\]M\. Bain, J\. Huh, T\. Han, and A\. Zisserman\(2023\)WhisperX: time\-accurate speech transcription of long\-form audio\.In24th Annual Conference of the International Speech Communication Association, Interspeech 2023, Dublin, Ireland, August 20\-24, 2023,N\. Harte, J\. Carson\-Berndsen, and G\. Jones \(Eds\.\),pp\. 4489–4493\.External Links:[Document](https://dx.doi.org/10.21437/INTERSPEECH.2023-78),[Link](https://doi.org/10.21437/Interspeech.2023-78)Cited by:[§II\-C](https://arxiv.org/html/2607.02633#S2.SS3.p1.1)\.
- \[4\]M\. Bernard and H\. Titeux\(2021\)Phonemizer: text to phones transcription for multiple languages in Python\.Journal of Open Source Software6\(68\),pp\. 3958\.Cited by:[§I](https://arxiv.org/html/2607.02633#S1.p3.1),[§II\-B](https://arxiv.org/html/2607.02633#S2.SS2.p1.1)\.
- \[5\]E\. Casanova, K\. Davis, E\. Gölge, G\. Göknar, I\. Gulea, L\. Hart, A\. Aljafari, J\. Meyer, R\. Morais, S\. Olayemi, and J\. Weber\(2024\)XTTS: a massively multilingual zero\-shot text\-to\-speech model\.InProc\. INTERSPEECH,Cited by:[§IV\-B](https://arxiv.org/html/2607.02633#S4.SS2.p1.1)\.
- \[6\]S\. Chen, C\. Wang, Z\. Chen,et al\.\(2022\)WavLM: large\-scale self\-supervised pre\-training for full stack speech processing\.IEEE Journal of Selected Topics in Signal Processing16\(6\),pp\. 1505–1518\.Cited by:[§IV\-C](https://arxiv.org/html/2607.02633#S4.SS3.p1.1)\.
- \[7\]Y\. Chen, Z\. Niu, Z\. Ma,et al\.\(2024\)F5\-TTS: a fairytaler that fakes fluent and faithful speech with flow matching\.ArXiv preprintabs/2410\.06885\.External Links:[Link](https://arxiv.org/abs/2410.06885)Cited by:[§IV\-B](https://arxiv.org/html/2607.02633#S4.SS2.p1.1)\.
- \[8\]H\. Choi, J\. Lee, W\. Kim, J\. Lee, H\. Heo, and K\. Lee\(2021\)Neural analysis and synthesis: reconstructing speech from self\-supervised representations\.InAdvances in Neural Information Processing Systems 34: Annual Conference on Neural Information Processing Systems 2021, NeurIPS 2021, December 6\-14, 2021, virtual,M\. Ranzato, A\. Beygelzimer, Y\. N\. Dauphin, P\. Liang, and J\. W\. Vaughan \(Eds\.\),pp\. 16251–16265\.External Links:[Link](https://proceedings.neurips.cc/paper/2021/hash/87682805257e619d49b8e0dfdc14affa-Abstract.html)Cited by:[§II\-D](https://arxiv.org/html/2607.02633#S2.SS4.p1.1)\.
- \[9\]B\. Desplanques, J\. Thienpondt, and K\. Demuynck\(2020\)ECAPA\-TDNN: emphasized channel attention, propagation and aggregation in TDNN based speaker verification\.InInterspeech 2020, 21st Annual Conference of the International Speech Communication Association, Virtual Event, Shanghai, China, 25\-29 October 2020,H\. Meng, B\. Xu, and T\. F\. Zheng \(Eds\.\),pp\. 3830–3834\.External Links:[Document](https://dx.doi.org/10.21437/Interspeech.2020-2650),[Link](https://doi.org/10.21437/Interspeech.2020-2650)Cited by:[§V\-A](https://arxiv.org/html/2607.02633#S5.SS1.p3.5)\.
- \[10\]Z\. Du, Q\. Chen, S\. Zhang,et al\.\(2024\)CosyVoice: a scalable multilingual zero\-shot text\-to\-speech synthesizer based on supervised semantic tokens\.ArXiv preprintabs/2407\.05407\.External Links:[Link](https://arxiv.org/abs/2407.05407)Cited by:[§I](https://arxiv.org/html/2607.02633#S1.p2.1),[§II\-A](https://arxiv.org/html/2607.02633#S2.SS1.p1.1)\.
- \[11\]Z\. Du, Y\. Wang, Q\. Chen,et al\.\(2024\)CosyVoice 2: scalable streaming speech synthesis with large language models\.ArXiv preprintabs/2412\.10117\.External Links:[Link](https://arxiv.org/abs/2412.10117)Cited by:[§IV\-B](https://arxiv.org/html/2607.02633#S4.SS2.p1.1)\.
- \[12\]J\. Duddingtonet al\.eSpeak NG text\-to\-speech\.Note:https://github\.com/espeak\-ng/espeak\-ngAccessed 2025Cited by:[§I](https://arxiv.org/html/2607.02633#S1.p3.1),[§II\-B](https://arxiv.org/html/2607.02633#S2.SS2.p1.1)\.
- \[13\]M\. Hansenet al\.Piper: a fast, local neural text\-to\-speech system\.Note:https://github\.com/rhasspy/piperAccessed 2025Cited by:[§IV\-B](https://arxiv.org/html/2607.02633#S4.SS2.p1.1)\.
- \[14\]M\. He, J\. Yang, L\. He, and F\. K\. Soong\(2022\)Neural lexicon reader: reduce pronunciation errors in end\-to\-end TTS by leveraging external textual knowledge\.In23rd Annual Conference of the International Speech Communication Association, Interspeech 2022, Incheon, Korea, September 18\-22, 2022,H\. Ko and J\. H\. L\. Hansen \(Eds\.\),pp\. 441–445\.External Links:[Document](https://dx.doi.org/10.21437/INTERSPEECH.2022-420),[Link](https://doi.org/10.21437/Interspeech.2022-420)Cited by:[§II\-B](https://arxiv.org/html/2607.02633#S2.SS2.p1.1),[§IV\-B](https://arxiv.org/html/2607.02633#S4.SS2.p1.1)\.
- \[15\]Y\. Jia, H\. Zen, J\. Shen, Y\. Zhang, and Y\. Wu\(2021\)PnG BERT: augmented BERT on phonemes and graphemes for neural TTS\.InInterspeech 2021, 22nd Annual Conference of the International Speech Communication Association, Brno, Czechia, 30 August \- 3 September 2021,H\. Hermansky, H\. Cernocký, L\. Burget, L\. Lamel, O\. Scharenborg, and P\. Motlícek \(Eds\.\),pp\. 151–155\.External Links:[Document](https://dx.doi.org/10.21437/Interspeech.2021-1757),[Link](https://doi.org/10.21437/Interspeech.2021-1757)Cited by:[§II\-B](https://arxiv.org/html/2607.02633#S2.SS2.p1.1)\.
- \[16\]F\. C\. JianStable\-ts: transcription, forced alignment, and audio indexing with OpenAI’s Whisper\.Note:https://github\.com/jianfch/stable\-tsAccessed 2026Cited by:[§II\-C](https://arxiv.org/html/2607.02633#S2.SS3.p1.1)\.
- \[17\]Z\. Ju, Y\. Wang, K\. Shen, X\. Tan, D\. Xin, D\. Yang, E\. Liu, Y\. Leng, K\. Song, S\. Tang, Z\. Wu, T\. Qin, X\. Li, W\. Ye, S\. Zhang, J\. Bian, L\. He, J\. Li, and S\. Zhao\(2024\)NaturalSpeech 3: zero\-shot speech synthesis with factorized codec and diffusion models\.InForty\-first International Conference on Machine Learning, ICML 2024, Vienna, Austria, July 21\-27, 2024,External Links:[Link](https://openreview.net/forum?id=dVhrnjZJad)Cited by:[§II\-A](https://arxiv.org/html/2607.02633#S2.SS1.p1.1),[§II\-B](https://arxiv.org/html/2607.02633#S2.SS2.p1.1)\.
- \[18\]E\. Kharitonov, D\. Vincent, Z\. Borsos, R\. Marinier, S\. Girgin, O\. Pietquin, M\. Sharifi, M\. Tagliasacchi, and N\. Zeghidour\(2023\)Speak, read and prompt: high\-fidelity text\-to\-speech with minimal supervision\.Transactions of the Association for Computational Linguistics11,pp\. 1703–1718\.External Links:[Document](https://dx.doi.org/10.1162/tacl%5Fa%5F00618),[Link](https://aclanthology.org/2023.tacl-1.95)Cited by:[§II\-A](https://arxiv.org/html/2607.02633#S2.SS1.p1.1)\.
- \[19\]J\. Kim, S\. Kim, J\. Kong, and S\. Yoon\(2020\)Glow\-tts: A generative flow for text\-to\-speech via monotonic alignment search\.InAdvances in Neural Information Processing Systems 33: Annual Conference on Neural Information Processing Systems 2020, NeurIPS 2020, December 6\-12, 2020, virtual,H\. Larochelle, M\. Ranzato, R\. Hadsell, M\. Balcan, and H\. Lin \(Eds\.\),External Links:[Link](https://proceedings.neurips.cc/paper/2020/hash/5c3b99e8f92532e5ad1556e53ceea00c-Abstract.html)Cited by:[§II\-B](https://arxiv.org/html/2607.02633#S2.SS2.p1.1)\.
- \[20\]J\. Kim, J\. Kong, and J\. Son\(2021\)Conditional variational autoencoder with adversarial learning for end\-to\-end text\-to\-speech\.InProceedings of the 38th International Conference on Machine Learning, ICML 2021, 18\-24 July 2021, Virtual Event,M\. Meila and T\. Zhang \(Eds\.\),Proceedings of Machine Learning Research, Vol\.139,pp\. 5530–5540\.External Links:[Link](http://proceedings.mlr.press/v139/kim21f.html)Cited by:[§II\-B](https://arxiv.org/html/2607.02633#S2.SS2.p1.1)\.
- \[21\]L\. Kürzinger, D\. Winkelbauer, L\. Li, T\. Watzel, and G\. Rigoll\(2020\)CTC\-segmentation of large corpora for German end\-to\-end speech recognition\.InProc\. SPECOM,pp\. 267–278\.Cited by:[§II\-C](https://arxiv.org/html/2607.02633#S2.SS3.p1.1)\.
- \[22\]M\. Le, A\. Vyas, B\. Shi, B\. Karrer, L\. Sari, R\. Moritz, M\. Williamson, V\. Manohar, Y\. Adi, J\. Mahadeokar, and W\. Hsu\(2023\)Voicebox: text\-guided multilingual universal speech generation at scale\.InAdvances in Neural Information Processing Systems 36: Annual Conference on Neural Information Processing Systems 2023, NeurIPS 2023, New Orleans, LA, USA, December 10 \- 16, 2023,A\. Oh, T\. Naumann, A\. Globerson, K\. Saenko, M\. Hardt, and S\. Levine \(Eds\.\),External Links:[Link](http://papers.nips.cc/paper%5C_files/paper/2023/hash/2d8911db9ecedf866015091b28946e15-Abstract-Conference.html)Cited by:[§II\-A](https://arxiv.org/html/2607.02633#S2.SS1.p1.1),[§II\-B](https://arxiv.org/html/2607.02633#S2.SS2.p1.1)\.
- \[23\]Y\. A\. Li, C\. Han, V\. S\. Raghavan, G\. Mischler, and N\. Mesgarani\(2023\)StyleTTS 2: towards human\-level text\-to\-speech through style diffusion and adversarial training with large speech language models\.InAdvances in Neural Information Processing Systems 36: Annual Conference on Neural Information Processing Systems 2023, NeurIPS 2023, New Orleans, LA, USA, December 10 \- 16, 2023,A\. Oh, T\. Naumann, A\. Globerson, K\. Saenko, M\. Hardt, and S\. Levine \(Eds\.\),External Links:[Link](http://papers.nips.cc/paper%5C_files/paper/2023/hash/3eaad2a0b62b5ed7a2e66c2188bb1449-Abstract-Conference.html)Cited by:[§IV\-B](https://arxiv.org/html/2607.02633#S4.SS2.p1.1)\.
- \[24\]S\. Liu\(2024\)Zero\-shot voice conversion with diffusion transformers\.ArXiv preprintabs/2411\.09943\.External Links:[Link](https://arxiv.org/abs/2411.09943)Cited by:[§II\-D](https://arxiv.org/html/2607.02633#S2.SS4.p1.1),[§III\-C](https://arxiv.org/html/2607.02633#S3.SS3.p2.11)\.
- \[25\]I\. Loshchilov and F\. Hutter\(2019\)Decoupled weight decay regularization\.In7th International Conference on Learning Representations, ICLR 2019, New Orleans, LA, USA, May 6\-9, 2019,External Links:[Link](https://openreview.net/forum?id=Bkg6RiCqY7)Cited by:[§III\-D](https://arxiv.org/html/2607.02633#S3.SS4.p1.6)\.
- \[26\]J\. LouradourWhisper\-timestamped\.Note:https://github\.com/linto\-ai/whisper\-timestampedAccessed 2026Cited by:[§II\-C](https://arxiv.org/html/2607.02633#S2.SS3.p1.1)\.
- \[27\]M\. McAuliffe, M\. Socolof, S\. Mihuc, M\. Wagner, and M\. Sonderegger\(2017\)Montreal forced aligner: trainable text\-speech alignment using kaldi\.InInterspeech 2017, 18th Annual Conference of the International Speech Communication Association, Stockholm, Sweden, August 20\-24, 2017,F\. Lacerda \(Ed\.\),pp\. 498–502\.External Links:[Link](http://www.isca-speech.org/archive/Interspeech%5C_2017/abstracts/1386.html)Cited by:[§II\-C](https://arxiv.org/html/2607.02633#S2.SS3.p1.1)\.
- \[28\]S\. Mehta, R\. Tu, J\. Beskow, É\. Székely, and G\. E\. Henter\(2024\)Matcha\-tts: A fast TTS architecture with conditional flow matching\.InIEEE International Conference on Acoustics, Speech and Signal Processing, ICASSP 2024, Seoul, Republic of Korea, April 14\-19, 2024,pp\. 11341–11345\.External Links:[Document](https://dx.doi.org/10.1109/ICASSP48485.2024.10448291),[Link](https://doi.org/10.1109/ICASSP48485.2024.10448291)Cited by:[§IV\-B](https://arxiv.org/html/2607.02633#S4.SS2.p1.1)\.
- \[29\]V\. Popov, I\. Vovk, V\. Gogoryan, T\. Sadekova, M\. S\. Kudinov, and J\. Wei\(2022\)Diffusion\-based voice conversion with fast maximum likelihood sampling scheme\.InThe Tenth International Conference on Learning Representations, ICLR 2022, Virtual Event, April 25\-29, 2022,External Links:[Link](https://openreview.net/forum?id=8c50f-DoWAu)Cited by:[§II\-D](https://arxiv.org/html/2607.02633#S2.SS4.p1.1)\.
- \[30\]V\. Pratap, Q\. Xu, A\. Sriram, G\. Synnaeve, and R\. Collobert\(2020\)MLS: A large\-scale multilingual dataset for speech research\.InInterspeech 2020, 21st Annual Conference of the International Speech Communication Association, Virtual Event, Shanghai, China, 25\-29 October 2020,H\. Meng, B\. Xu, and T\. F\. Zheng \(Eds\.\),pp\. 2757–2761\.External Links:[Document](https://dx.doi.org/10.21437/Interspeech.2020-2826),[Link](https://doi.org/10.21437/Interspeech.2020-2826)Cited by:[TABLE I](https://arxiv.org/html/2607.02633#S3.T1)\.
- \[31\]K\. Qian, Y\. Zhang, S\. Chang, X\. Yang, and M\. Hasegawa\-Johnson\(2019\)AutoVC: zero\-shot voice style transfer with only autoencoder loss\.InProceedings of the 36th International Conference on Machine Learning, ICML 2019, 9\-15 June 2019, Long Beach, California, USA,K\. Chaudhuri and R\. Salakhutdinov \(Eds\.\),Proceedings of Machine Learning Research, Vol\.97,pp\. 5210–5219\.External Links:[Link](http://proceedings.mlr.press/v97/qian19c.html)Cited by:[§II\-D](https://arxiv.org/html/2607.02633#S2.SS4.p1.1)\.
- \[32\]K\. Qian, Y\. Zhang, H\. Gao, J\. Ni, C\. Lai, D\. D\. Cox, M\. Hasegawa\-Johnson, and S\. Chang\(2022\)ContentVec: an improved self\-supervised speech representation by disentangling speakers\.InInternational Conference on Machine Learning, ICML 2022, 17\-23 July 2022, Baltimore, Maryland, USA,K\. Chaudhuri, S\. Jegelka, L\. Song, C\. Szepesvári, G\. Niu, and S\. Sabato \(Eds\.\),Proceedings of Machine Learning Research, Vol\.162,pp\. 18003–18017\.External Links:[Link](https://proceedings.mlr.press/v162/qian22b.html)Cited by:[§II\-D](https://arxiv.org/html/2607.02633#S2.SS4.p1.1)\.
- \[33\]Qwen Team\(2026\)Qwen3\-TTS technical report\.ArXiv preprintabs/2601\.15621\.External Links:[Link](https://arxiv.org/abs/2601.15621)Cited by:[§I](https://arxiv.org/html/2607.02633#S1.p2.1),[§II\-A](https://arxiv.org/html/2607.02633#S2.SS1.p1.1),[§II\-C](https://arxiv.org/html/2607.02633#S2.SS3.p1.1),[§III\-A](https://arxiv.org/html/2607.02633#S3.SS1.p1.12),[§IV\-B](https://arxiv.org/html/2607.02633#S4.SS2.p1.1)\.
- \[34\]A\. Radford, J\. W\. Kim, T\. Xu, G\. Brockman, C\. McLeavey, and I\. Sutskever\(2023\)Robust speech recognition via large\-scale weak supervision\.InInternational Conference on Machine Learning, ICML 2023, 23\-29 July 2023, Honolulu, Hawaii, USA,A\. Krause, E\. Brunskill, K\. Cho, B\. Engelhardt, S\. Sabato, and J\. Scarlett \(Eds\.\),Proceedings of Machine Learning Research, Vol\.202,pp\. 28492–28518\.External Links:[Link](https://proceedings.mlr.press/v202/radford23a.html)Cited by:[§IV\-C](https://arxiv.org/html/2607.02633#S4.SS3.p1.1),[§V\-A](https://arxiv.org/html/2607.02633#S5.SS1.p2.16)\.
- \[35\]Y\. Ren, C\. Hu, X\. Tan, T\. Qin, S\. Zhao, Z\. Zhao, and T\. Liu\(2021\)FastSpeech 2: fast and high\-quality end\-to\-end text to speech\.In9th International Conference on Learning Representations, ICLR 2021, Virtual Event, Austria, May 3\-7, 2021,External Links:[Link](https://openreview.net/forum?id=piLPYqxtWuA)Cited by:[§I](https://arxiv.org/html/2607.02633#S1.p1.1),[§II\-B](https://arxiv.org/html/2607.02633#S2.SS2.p1.1)\.
- \[36\]T\. Saeki, D\. Xin, W\. Nakata, T\. Koriyama, S\. Takamichi, and H\. Saruwatari\(2022\)UTMOS: utokyo\-sarulab system for voicemos challenge 2022\.In23rd Annual Conference of the International Speech Communication Association, Interspeech 2022, Incheon, Korea, September 18\-22, 2022,H\. Ko and J\. H\. L\. Hansen \(Eds\.\),pp\. 4521–4525\.External Links:[Document](https://dx.doi.org/10.21437/INTERSPEECH.2022-439),[Link](https://doi.org/10.21437/Interspeech.2022-439)Cited by:[§IV\-C](https://arxiv.org/html/2607.02633#S4.SS3.p1.1)\.
- \[37\]E\. Salesky, K\. Darwish, M\. Al\-Badrashiny, M\. Diab, and J\. Niehues\(2023\)Evaluating multilingual speech translation under realistic conditions with resegmentation and terminology\.InProceedings of the 20th International Conference on Spoken Language Translation \(IWSLT 2023\),E\. Salesky, M\. Federico, and M\. Carpuat \(Eds\.\),Toronto, Canada \(in\-person and online\),pp\. 62–78\.External Links:[Document](https://dx.doi.org/10.18653/v1/2023.iwslt-1.2),[Link](https://aclanthology.org/2023.iwslt-1.2)Cited by:[TABLE II](https://arxiv.org/html/2607.02633#S3.T2)\.
- \[38\]J\. Shen, R\. Pang, R\. J\. Weiss, M\. Schuster, N\. Jaitly, Z\. Yang, Z\. Chen, Y\. Zhang, Y\. Wang, R\. Ryan, R\. A\. Saurous, Y\. Agiomyrgiannakis, and Y\. Wu\(2018\)Natural TTS synthesis by conditioning wavenet on MEL spectrogram predictions\.In2018 IEEE International Conference on Acoustics, Speech and Signal Processing, ICASSP 2018, Calgary, AB, Canada, April 15\-20, 2018,pp\. 4779–4783\.External Links:[Document](https://dx.doi.org/10.1109/ICASSP.2018.8461368),[Link](https://doi.org/10.1109/ICASSP.2018.8461368)Cited by:[§I](https://arxiv.org/html/2607.02633#S1.p1.1),[§II\-B](https://arxiv.org/html/2607.02633#S2.SS2.p1.1)\.
- \[39\]K\. Shen, Z\. Ju, X\. Tan, E\. Liu, Y\. Leng, L\. He, T\. Qin, S\. Zhao, and J\. Bian\(2024\)NaturalSpeech 2: latent diffusion models are natural and zero\-shot speech and singing synthesizers\.InThe Twelfth International Conference on Learning Representations, ICLR 2024, Vienna, Austria, May 7\-11, 2024,External Links:[Link](https://openreview.net/forum?id=Rc7dAwVL3v)Cited by:[§II\-A](https://arxiv.org/html/2607.02633#S2.SS1.p1.1)\.
- \[40\]Silero TeamSilero VAD: pre\-trained enterprise\-grade voice activity detector\.Note:https://github\.com/snakers4/silero\-vadAccessed 2025Cited by:[§III\-E](https://arxiv.org/html/2607.02633#S3.SS5.p1.4)\.
- \[41\]A\. Singhet al\.\(2026\)SonoEdit: null\-space constrained knowledge editing for pronunciation correction in LLM\-based TTS\.ArXiv preprintabs/2601\.17086\.External Links:[Link](https://arxiv.org/abs/2601.17086)Cited by:[§II\-B](https://arxiv.org/html/2607.02633#S2.SS2.p1.1),[§IV\-B](https://arxiv.org/html/2607.02633#S4.SS2.p1.1)\.
- \[42\]C\. Wang, S\. Chen, Y\. Wu,et al\.\(2023\)Neural codec language models are zero\-shot text to speech synthesizers\.ArXiv preprintabs/2301\.02111\.External Links:[Link](https://arxiv.org/abs/2301.02111)Cited by:[§I](https://arxiv.org/html/2607.02633#S1.p2.1),[§II\-A](https://arxiv.org/html/2607.02633#S2.SS1.p1.1)\.
- \[43\]T\. Wang, H\. Wang, M\. Ge, C\. Gong, C\. Qiang, Z\. Ma, Z\. Huang, G\. Yang, X\. Wang, E\. S\. Chng, X\. Chen, L\. Wang, and J\. Dang\(2025\)Word\-level emotional expression control in zero\-shot text\-to\-speech synthesis\.InAdvances in Neural Information Processing Systems 38 \(NeurIPS 2025\),Cited by:[§II\-A](https://arxiv.org/html/2607.02633#S2.SS1.p1.1)\.
- \[44\]Y\. Wang, H\. Zhan, L\. Liu,et al\.\(2025\)MaskGCT: zero\-shot text\-to\-speech with masked generative codec transformer\.InProc\. ICLR,Cited by:[§II\-A](https://arxiv.org/html/2607.02633#S2.SS1.p1.1),[§IV\-B](https://arxiv.org/html/2607.02633#S4.SS2.p1.1)\.
- \[45\]Q\. Xu, A\. Baevski, and M\. Auli\(2022\)Simple and effective zero\-shot cross\-lingual phoneme recognition\.In23rd Annual Conference of the International Speech Communication Association, Interspeech 2022, Incheon, Korea, September 18\-22, 2022,H\. Ko and J\. H\. L\. Hansen \(Eds\.\),pp\. 2113–2117\.External Links:[Document](https://dx.doi.org/10.21437/INTERSPEECH.2022-60),[Link](https://doi.org/10.21437/Interspeech.2022-60)Cited by:[§I](https://arxiv.org/html/2607.02633#S1.p3.1),[TABLE III](https://arxiv.org/html/2607.02633#S3.T3),[§IV\-B](https://arxiv.org/html/2607.02633#S4.SS2.p1.1)\.
- \[46\]J\. Yamagishi, C\. Veaux, and K\. MacDonald\(2019\)CSTR VCTK corpus: English multi\-speaker corpus for CSTR voice cloning toolkit\.Technical reportUniversity of Edinburgh, CSTR\.Cited by:[TABLE I](https://arxiv.org/html/2607.02633#S3.T1)\.
- \[47\]Z\. Ye, X\. Zhu, C\.\-M\. Chan,et al\.\(2025\)Llasa: scaling train\-time and inference\-time compute for Llama\-based speech synthesis\.ArXiv preprintabs/2502\.04128\.External Links:[Link](https://arxiv.org/abs/2502.04128)Cited by:[§I](https://arxiv.org/html/2607.02633#S1.p2.1),[§II\-A](https://arxiv.org/html/2607.02633#S2.SS1.p1.1)\.
- \[48\]H\. Zen, V\. Dang, R\. Clark, Y\. Zhang, R\. J\. Weiss, Y\. Jia, Z\. Chen, and Y\. Wu\(2019\)LibriTTS: A corpus derived from librispeech for text\-to\-speech\.InInterspeech 2019, 20th Annual Conference of the International Speech Communication Association, Graz, Austria, 15\-19 September 2019,G\. Kubin and Z\. Kacic \(Eds\.\),pp\. 1526–1530\.External Links:[Document](https://dx.doi.org/10.21437/Interspeech.2019-2441),[Link](https://doi.org/10.21437/Interspeech.2019-2441)Cited by:[TABLE I](https://arxiv.org/html/2607.02633#S3.T1)\.
- \[49\]J\. Zhu, F\. Samir, E\. Chodroff, and D\. R\. Mortensen\(2025\)ZIPA: a family of efficient models for multilingual phone recognition\.InProc\. ACL,pp\. 19568–19585\.Cited by:[§I](https://arxiv.org/html/2607.02633#S1.p3.1),[§IV\-C](https://arxiv.org/html/2607.02633#S4.SS3.p1.1)\.
- \[50\]J\. Zhu, C\. Zhang, and D\. Jurgens\(2022\)ByT5 model for massively multilingual grapheme\-to\-phoneme conversion\.In23rd Annual Conference of the International Speech Communication Association, Interspeech 2022, Incheon, Korea, September 18\-22, 2022,H\. Ko and J\. H\. L\. Hansen \(Eds\.\),pp\. 446–450\.External Links:[Document](https://dx.doi.org/10.21437/INTERSPEECH.2022-538),[Link](https://doi.org/10.21437/Interspeech.2022-538)Cited by:[§II\-B](https://arxiv.org/html/2607.02633#S2.SS2.p1.1)\.