DLLM-TTS: Block Discrete Diffusion Language Model for Text-to-Speech Synthesis

# DLLM-TTS: Block Discrete Diffusion Language Model for Text-to-Speech Synthesis
Source: [https://arxiv.org/html/2608.00011](https://arxiv.org/html/2608.00011)
Madha Mathur Sayed Singh Khurana Mandloi Kamath

###### Abstract

Current text\-to\-speech systems face a trade\-off: autoregressive codec language models produce highly intelligible speech but require large\-scale models and training data and decode tokens sequentially, while non\-autoregressive approaches improve speed at the cost of linguistic accuracy\. We present DLLM\-TTS, a framework that formulates TTS as conditional block discrete diffusion over X\-Codec2 neural audio codec tokens\. The model decomposes sequences into blocks and applies masked diffusion within each block while processing blocks sequentially, learning both local acoustic coherence and global text\-speech alignment\. During inference, parallel token prediction within blocks enables efficient generation with a real\-time factor \(RTF\) of 0\.15\. A 0\.6B\-parameter model trained on 20K hours achieves competitive performance on the Seed\-TTS\-eval benchmark, demonstrating that block discrete diffusion language models enable practical and data\-efficient speech synthesis with parallel generation\.

###### keywords:

discrete diffusion, text\-to\-speech, block diffusion, neural codec, language model

## 1Introduction

Text\-to\-speech synthesis faces a fundamental efficiency\-quality trade\-off\. Autoregressive codec language models\[wang2023valle,chen2024valle2,ye2025llasa\]achieve high\-quality zero\-shot synthesis but require 60K–250K hours of data and generate tokens sequentially, incurring high latency\. Non\-autoregressive approaches based on flow matching\[le2024voicebox,chen2024f5tts\]and diffusion\[ju2024naturalspeech3,li2023styletts2\]enable parallel generation but typically need explicit duration modeling or struggle with text\-speech alignment, leading to word skipping or repetition errors\.

Recent masked discrete diffusion language models\[sahoo2024mdlm,nie2025llada\]have shown that discrete diffusion can match autoregressive quality for text generation with parallel decoding\. BD3\-LM\[arriola2025bd3lm\]further introduces block decomposition, interpolating between autoregressive and fully parallel generation with flexible speed\-quality trade\-offs\. However, these methods have only been explored for text; their application to conditional speech generation over discrete codec tokens remains unexplored\.

We presentDLLM\-TTS, which adapts block discrete diffusion to conditional speech generation\. Speech exhibits strong local acoustic coherence at the phoneme level while requiring longer\-range dependencies for text\-speech alignment, a structure naturally suited to block diffusion\. We model speech as block discrete diffusion over X\-Codec2\[ye2024xcodec\]codec tokens: sequences are decomposed into blocks, and the model reconstructs masked tokens within each block conditioned on text, enabling parallel prediction within blocks while preserving sequential dependencies across them\. The masked diffusion objective additionally provides implicit data augmentation, as each sequence is observed under diverse masking patterns, improving data efficiency over autoregressive training\.

Contributions\.❶ We introduce block discrete diffusion for TTS, modeling speech synthesis as conditional masked token reconstruction over codec tokens with block decomposition, the first application of block discrete diffusion to conditional speech generation\. ❷ We employ staircase attention to jointly capture local acoustic coherence within blocks and global text\-speech alignment across blocks, without explicit duration modeling\. ❸ We achieve competitive intelligibility on the Seed\-TTS benchmark\[anastassiou2024seedtts\]using only 20K hours of data, a 3–12×\\timesreduction over autoregressive systems\. ❹ A 0\.6B\-parameter model achieves an RTF of 0\.15 with strong zero\-shot speaker similarity\.

## 2Related Work

Autoregressive Codec Language Models\.A dominant paradigm in TTS formulates speech synthesis as language modeling over discrete neural codec tokens\. VALL\-E\[wang2023valle\]pioneered this approach by training on 60K hours of speech, achieving strong zero\-shot synthesis from a 3\-second prompt\. VALL\-E 2\[chen2024valle2\]introduced repetition\-aware sampling and grouped code modeling to reach human parity on standard benchmarks\. LLASA\[ye2025llasa\]demonstrated that scaling LLaMA\-based architectures with X\-Codec2\[ye2024xcodec\]up to 8B parameters yields consistent quality gains\. DiTAR\[jia2025ditar\]combines a language model with a diffusion transformer in a patch\-based autoregressive framework for continuous\-valued speech generation\. While these systems achieve high quality, sequential token generation introduces latency that limits real\-time applications\.

Non\-Autoregressive Speech Synthesis\.To overcome the latency bottleneck, non\-autoregressive methods generate speech in parallel\. Voicebox\[le2024voicebox\]applies flow matching over mel\-spectrograms conditioned on text and surrounding audio context\. F5\-TTS\[chen2024f5tts\]simplifies the pipeline by using flow matching with a Diffusion Transformer \(DiT\), eliminating duration models and phoneme alignment\. NaturalSpeech 3\[ju2024naturalspeech3\]employs factorized diffusion over disentangled speech attributes \(content, prosody, timbre\)\. StyleTTS 2\[li2023styletts2\]achieves human\-level quality through style diffusion and adversarial training with speech language model discriminators\. CosyVoice\[du2024cosyvoice\]combines an LLM\-based text\-to\-token stage with conditional flow matching for token\-to\-speech synthesis\. SoundStorm\[borsos2023soundstorm\]uses confidence\-based parallel decoding over RVQ tokens for fast generation but struggles with text\-speech alignment when trained from scratch\. MegaTTS 3\[jiang2025megatts3\]proposes sparse alignment with a latent diffusion transformer for robust zero\-shot synthesis\. These approaches typically require explicit duration prediction, frame\-aligned annotations, or multi\-stage pipelines, which complicate training data preparation\.

Discrete Diffusion Language Models\.Masked discrete diffusion has recently emerged as a viable alternative to autoregressive modeling for text generation\. MDLM\[sahoo2024mdlm\]derives a simplified objective based on masked language modeling losses, achieving strong performance among diffusion\-based language models\. LLaDA\[nie2025llada\]scales masked diffusion to 8B parameters, demonstrating competitive performance with autoregressive LLMs on in\-context learning and instruction following\. BD3\-LM\[arriola2025bd3lm\]extends MDLM with block decomposition, interpolating between fully parallel diffusion and autoregressive generation while enabling variable\-length output and KV caching\. However, these methods have primarily been studied for text generation\. Their application to conditional speech generation, where both acoustic coherence and text–speech alignment must be modeled jointly over discrete codec tokens, remains largely unexplored\. Our work addresses this gap by adapting block discrete diffusion for text\-to\-speech synthesis\.

## 3Methodology

Input Token Sequenceref\_txtref\_codecgen\_txt\[MASK\]L\[\\text\{MASK\}\]^\{L\}Block Discrete Diffusion TransformerIterative DenoisingBlk 1Blk kBlk NX\-Codec2 DecoderWaveform

Figure 1:Overview of DLLM\-TTS\.Textandreference codec tokensare concatenated withgeneration textand\[MASK\]Ltargets, then processed by theBlock Discrete Diffusion Transformerusing staircase attention \(bidirectional/causal\)\. Each block isiteratively denoisedand decoded byX\-Codec2into speech\.### 3\.1Background

Neural Audio Codecs\.Neural audio codecs compress continuous audio waveforms into discrete token sequences through learned quantization\. Traditional residual vector quantization \(RVQ\) approaches like EnCodec\[defossez2022encodec\]employ multiple codebook layers to progressively refine representations, producing parallel token streams\. In contrast, X\-Codec2\[ye2024xcodec\]adopts a unified semantic\-acoustic architecture: a semantic encoder \(Wav2Vec2\-BERT\) captures linguistic content while an acoustic encoder preserves fine\-grained audio characteristics\. These representations are fused and quantized using single\-stage Finite Scalar Quantization \(FSQ\) with vocabulary size\|𝒱\|=6561\|\\mathcal\{V\}\|=6561, producing a single token stream at frame ratefr=50f\_\{r\}=50Hz\. This yieldsL=fr⋅dsecL=f\_\{r\}\\cdot d\_\{\\text\{sec\}\}codec tokens fordsecd\_\{\\text\{sec\}\}seconds of audio, simplifying integration with transformer language models by avoiding multiple parallel codebook streams\.

Masked Discrete Diffusion\.Masked diffusion\[sahoo2024mdlm\]defines a forward process that progressively corrupts discrete sequences by replacing tokens with a special\[MASK\]token, and a reverse process that learns to denoise the corrupted observations\. For a sequence𝐱=\(x1,…,xL\)\\mathbf\{x\}=\(x\_\{1\},\\ldots,x\_\{L\}\)where eachxi∈𝒱x\_\{i\}\\in\\mathcal\{V\}, the forward process at continuous timestept∈\[0,1\]t\\in\[0,1\]corrupts each token independently:

q​\(𝐳t∣𝐱\)=∏i=1L\[αt⋅δ​\(zti=xi\)\+\(1−αt\)⋅δ​\(zti=\[MASK\]\)\]q\(\\mathbf\{z\}\_\{t\}\\mid\\mathbf\{x\}\)=\\prod\_\{i=1\}^\{L\}\\big\[\\alpha\_\{t\}\\cdot\\delta\(z\_\{t\}^\{i\}=x\_\{i\}\)\+\(1\-\\alpha\_\{t\}\)\\cdot\\delta\(z\_\{t\}^\{i\}=\\texttt\{\[MASK\]\}\)\\big\]\(1\)whereαt∈\[0,1\]\\alpha\_\{t\}\\in\[0,1\]is a monotonically decreasing schedule function\. Att=0t=0the sequence is fully clean \(α0=1\\alpha\_\{0\}=1\), while att=1t=1it is fully masked \(α1≈0\\alpha\_\{1\}\\approx 0\)\. The model learns the reverse processpθ​\(𝐱∣𝐳t,𝐜\)p\_\{\\theta\}\(\\mathbf\{x\}\\mid\\mathbf\{z\}\_\{t\},\\mathbf\{c\}\)to predict original tokens given corrupted observations𝐳t\\mathbf\{z\}\_\{t\}and conditioning𝐜\\mathbf\{c\}\.

Block Discrete Diffusion\.BD3\-LM\[arriola2025bd3lm\]extends masked diffusion by decomposing sequences into blocks of sizeBB\. Rather than processing the entire sequence uniformly, block diffusion generates blocks sequentially while allowing parallel prediction within each block\. During training, the model receives concatenated input\[𝐱t⊕𝐱0\]\[\\mathbf\{x\}\_\{t\}\\oplus\\mathbf\{x\}\_\{0\}\]where𝐱t\\mathbf\{x\}\_\{t\}contains masked tokens and𝐱0\\mathbf\{x\}\_\{0\}contains clean tokens from previous blocks\. A specializedstaircase attentionmask enforces: \(1\) bidirectional attention within each block for local coherence, \(2\) causal attention from noised blocks to previous clean blocks for sequential context, and \(3\) full causal attention within clean blocks\. This architecture interpolates between fully parallel diffusion \(B=LB=L\) and autoregressive generation \(B=1B=1\), enabling a flexible trade\-off between generation speed and sequential dependency modeling\.

### 3\.2Block Discrete Diffusion for Speech

We formulate text\-to\-speech synthesis as conditional block discrete diffusion over codec token sequences\. Given a text input and speaker prompt, the model generates speech by iteratively denoising masked codec tokens through a block\-based generation process\.

Sequence Decomposition\.A codec token sequence𝐱=\(x1,…,xL\)\\mathbf\{x\}=\(x\_\{1\},\\ldots,x\_\{L\}\)is partitioned intoKKcontiguous blocks of sizeBB:

𝐱=\[𝐱\(1\),…,𝐱\(K\)\],K=⌈L/B⌉,𝐱\(k\)∈𝒱B\\mathbf\{x\}=\[\\mathbf\{x\}^\{\(1\)\},\\ldots,\\mathbf\{x\}^\{\(K\)\}\],\\quad K=\\lceil L/B\\rceil,\\quad\\mathbf\{x\}^\{\(k\)\}\\in\\mathcal\{V\}^\{B\}\(2\)whereB=32B=32tokens \(∼\\sim0\.64 s of audio atfr=50f\_\{r\}=50Hz\)\.

Forward Process\.For a given diffusion timestept∈\[0,1\]t\\in\[0,1\], tokens within the target block are independently masked with probability1−αt1\-\\alpha\_\{t\}whereαt=1−t\\alpha\_\{t\}=1\-tis a linear schedule\. Concretely, for each positioniiin blockkk, a binary mask is sampled asmi\(k\)∼Bernoulli​\(1−αt\)m\_\{i\}^\{\(k\)\}\\sim\\text\{Bernoulli\}\(1\-\\alpha\_\{t\}\), and the corrupted token is:

zti=\(1−mi\(k\)\)​xi\+mi\(k\)​\[MASK\]z\_\{t\}^\{i\}=\(1\-m\_\{i\}^\{\(k\)\}\)\\,x\_\{i\}\+m\_\{i\}^\{\(k\)\}\\,\\texttt\{\[MASK\]\}\(3\)i\.e\., positions wheremi\(k\)=1m\_\{i\}^\{\(k\)\}=1are replaced with\[MASK\]while others remain clean\. The resulting masked sequence𝐳t\\mathbf\{z\}\_\{t\}forms the corrupted observation that the model learns to reconstruct\.

Reverse Process and Training\.The model predicts original unmasked tokens given𝐳t\\mathbf\{z\}\_\{t\}, conditioned on both text and speaker information\. We adopt staircase attention, where positioniiattends to positionjjaccording to the binary mask𝐀i​j∈\{0,1\}\\mathbf\{A\}\_\{ij\}\\in\\\{0,1\\\}, defined as:

𝐀i​j=𝟏​\[C1∨C2∨C3\]\\mathbf\{A\}\_\{ij\}=\\mathbf\{1\}\[\\,C\_\{1\}\\vee C\_\{2\}\\vee C\_\{3\}\\,\]\(4\)with three conditions:C1:C\_\{1\}\{:\}bi=bj∧bi≤b∗\\;b\_\{i\}\{=\}b\_\{j\}\\wedge b\_\{i\}\{\\leq\}b^\{\*\}\(bidirectional within noised blocks\),C2:bj<biC\_\{2\}\{:\}\\;b\_\{j\}\{<\}b\_\{i\}\(causal across blocks\), andC3:bi=bj∧bi\>b∗∧j≤iC\_\{3\}\{:\}\\;b\_\{i\}\{=\}b\_\{j\}\\wedge b\_\{i\}\{\>\}b^\{\*\}\\wedge j\{\\leq\}i\(causal within clean blocks\), wherebi=⌈i/B⌉b\_\{i\}=\\lceil i/B\\rceilmaps positioniito its block index,b∗b^\{\*\}is the current noised block, and𝟏​\[⋅\]\\mathbf\{1\}\[\\cdot\]is the indicator function\. This enables bidirectional attention within each noised block for local coherence, causal attention to previous clean blocks for sequential context, and causal attention within clean blocks for autoregressive conditioning\. The training objective minimizes cross\-entropy loss over masked positions:

ℒ=𝔼t∼𝒰\[0,1\],𝐱,𝐳t∼q\(⋅\|𝐱\)​\[−∑i∈ℳtlog⁡pθ​\(xi∣𝐳t,𝐜\)\]\\mathcal\{L\}=\\mathbb\{E\}\_\{t\\sim\\mathcal\{U\}\[0,1\],\\,\\mathbf\{x\},\\,\\mathbf\{z\}\_\{t\}\\sim q\(\\cdot\|\\mathbf\{x\}\)\}\\bigg\[\-\\sum\_\{i\\in\\mathcal\{M\}\_\{t\}\}\\log p\_\{\\theta\}\(x\_\{i\}\\mid\\mathbf\{z\}\_\{t\},\\mathbf\{c\}\)\\bigg\]\(5\)whereℳt\\mathcal\{M\}\_\{t\}denotes the set of masked positions at timesteptt, and𝐜\\mathbf\{c\}represents conditioning information \(text and speaker prompt\)\.

Variable\-Length Handling\.To handle variable\-length sequences, we introduce an end\-of\-sequence \(EOS\) token\. All tokens following EOS are set to EOS, which simplifies the masking process and eliminates the need for explicit length prediction, as the model learns to generate EOS when synthesis is complete\.

### 3\.3Model Architecture

Our model is a 0\.6B\-parameter transformer initialized from Qwen2\[yang2024qwen2\]and adapted for block discrete diffusion\. Table[1](https://arxiv.org/html/2608.00011#S3.T1)summarizes the architecture\. The model uses Rotary Position Embeddings \(RoPE\)\[su2024rope\]for efficient long\-context modeling\.

Table 1:Model architecture hyperparameters\.Training Input Format\.During training, the input sequence consists ofNNtext tokens followed byLLcodec tokens:

\[<text\>​t1​⋯​tN​<EOS\>​c1​⋯​cL\]\[\\texttt\{<text\>\}\\,t\_\{1\}\\cdots t\_\{N\}\\,\\texttt\{<EOS\>\}\\,c\_\{1\}\\cdots c\_\{L\}\]Text tokens and codec tokens are embedded via separate embedding layers𝐄text∈ℝ\|𝒱text\|×d\\mathbf\{E\}\_\{\\text\{text\}\}\\in\\mathbb\{R\}^\{\|\\mathcal\{V\}\_\{\\text\{text\}\}\|\\times d\}and𝐄codec∈ℝ\|𝒱\|×d\\mathbf\{E\}\_\{\\text\{codec\}\}\\in\\mathbb\{R\}^\{\|\\mathcal\{V\}\|\\times d\}into a shareddd\-dimensional space, then concatenated to form the input𝐡0∈ℝ\(N\+L\)×d\\mathbf\{h\}\_\{0\}\\in\\mathbb\{R\}^\{\(N\+L\)\\times d\}\. During each training step, codec tokens are randomly masked according to the diffusion timesteptt\(Eq\.[3](https://arxiv.org/html/2608.00011#S3.E3)\), and the model learns to reconstruct the original tokens\.

Inference Input Format\.For zero\-shot speaker adaptation, we employ a reference\-and\-generation paradigm:

The reference text is the transcript of a 3–5 s speaker prompt\. The reference codec tokens \(M=150M=150–250250atfr=50f\_\{r\}=50Hz\) encode the speaker's voice characteristics\. The generation section \(LLtokens\) starts fully masked and is iteratively denoised through block diffusion \(Eq\.[2](https://arxiv.org/html/2608.00011#S3.E2)\), providing the model with both linguistic content and speaker identity through attention\-based conditioning\. The maximum sequence length is 2048 tokens \(∼\\sim40 s of speech\)\.

### 3\.4Inference

At inference time, we generate speech through sequential block diffusion decoding\. Starting from a fully masked generation segment, we processKKblocks sequentially, applyingTTdenoising steps within each block \(defaultT=B/2=16T=B/2=16\)\.

Confidence\-Based Sampling\.At each denoising steps∈\{1,…,T\}s\\in\\\{1,\\ldots,T\\\}within a block, the model predicts token distributions for all masked positions\. For each masked positioni∈ℳsi\\in\\mathcal\{M\}\_\{s\}, we compute:

x^i=arg⁡maxv∈𝒱⁡pθ​\(xi=v∣𝐳s,𝐜\)\\hat\{x\}\_\{i\}=\\arg\\max\_\{v\\in\\mathcal\{V\}\}p\_\{\\theta\}\(x\_\{i\}=v\\mid\\mathbf\{z\}\_\{s\},\\mathbf\{c\}\)\(6\)Positioniiis unmasked if the model confidence exceeds a thresholdτ\\tau:

𝒰s=\{i∈ℳs:maxv⁡pθ​\(xi=v∣𝐳s,𝐜\)\>τ\}\\mathcal\{U\}\_\{s\}=\\big\\\{i\\in\\mathcal\{M\}\_\{s\}:\\max\_\{v\}p\_\{\\theta\}\(x\_\{i\}=v\\mid\\mathbf\{z\}\_\{s\},\\mathbf\{c\}\)\>\\tau\\big\\\}\(7\)withτ=0\.6\\tau=0\.6\. Unmasked positions are fixed tox^i\\hat\{x\}\_\{i\}; remaining positions stay masked for subsequent steps\. This allows the model to commit to high\-certainty tokens first, then resolve ambiguous positions\.

Early Stopping\.If all masked positions are unmasked \(ℳs=∅\\mathcal\{M\}\_\{s\}=\\emptyset\) before stepTT, denoising terminates early, reducing computation for easy blocks\.

Generation Speed\.Each block ofB=32B=32tokens spansB/fr=0\.64B/f\_\{r\}=0\.64s of audio\. Blocks are decoded sequentially with up toTTparallel denoising steps each, yielding an RTF of 0\.15 atT=16T=16\. KV caching across blocks reduces latency for subsequent blocks, and the block\-sequential design enables streaming with low time\-to\-first\-audio\.

## 4Experiments

### 4\.1Experimental Setup

Training Data\.We train DLLM\-TTS through a two\-stage curriculum:Stage 1trains the model on 16K hours sampled from the Emilia dataset\[he2024emilia\]for 20 epochs, establishing coherent codec token generation conditioned on text and speaker prompts\. Training uses a batch size of 16 per GPU with 8\-step gradient accumulation \(effective batch size 128\) across8×8\\timesH100 GPUs for 3 days\.Stage 2fine\-tunes on 4K hours of high\-quality synthetic speech to improve prosody and alignment\. Throughout both stages, we samplet∼𝒰​\[0,1\]t\\sim\\mathcal\{U\}\[0,1\]and apply per\-token masking with probability1−αt1\-\\alpha\_\{t\}\(Eq\.[3](https://arxiv.org/html/2608.00011#S3.E3)\)\. We use AdamW\[loshchilov2019adamw\]with learning rate1×10−41\\times 10^\{\-4\}, cosine schedule with 1% warmup, and effective batch size 128\.

Evaluation\.We evaluate on theSeed\-TTS\-evalbenchmark\[anastassiou2024seedtts\], a zero\-shot TTS evaluation suite covering standard and challenging scenarios \(rare words, complex prosody, long\-form utterances\)\.

Metrics\.We report:

- •Word Error Rate \(WER\)andCharacter Error Rate \(CER\)computed using Whisper\-large\-v3\[radford2023whisper\]to measure intelligibility\.
- •Speaker Similarity \(SIM\)measured as cosine similarity between speaker embeddings extracted using WavLM\-TDNN\[chen2022wavlm\], evaluating voice cloning fidelity\.
- •Mean Opinion Score \(MOS\)collected from 25 listeners following the CodecMOS\-Accent protocol\[huang2026codecmos\], assessing perceptual naturalness and prosody on a 5\-point scale\.

Baselines\.We compare against autoregressive models \(LLASA\[ye2025llasa\], IndexTTS2\[zhou2025indextts2\], Qwen2\.5\-Omni\[xu2025qwenomni\]\), non\-autoregressive models \(F5\-TTS\[chen2024f5tts\], MaskGCT\[wang2024maskgct\], CosyVoice3\[du2025cosyvoice3\], OpenAudio\-s1\-mini\[openaudio2024\]\), and hybrid models \(DiTAR\[jia2025ditar\]\)\.

### 4\.2Results and Analysis

Table[2](https://arxiv.org/html/2608.00011#S4.T2)presents results on the Seed\-TTS\-eval benchmark\. WithT=32T=32denoising steps and block sizeB=32B=32,DLLM\-TTS achieves a WER of 2\.25% and one of the highest speaker similarity scores \(0\.750\) among open\-source systems, using only0\.6B parametersand20K hours of training data, substantially less than systems trained on up to 250K hours\. These results demonstrate that block discrete diffusion achieves competitive intelligibility while providing strong zero\-shot voice cloning\.

Subjective Quality\.On subjective evaluation \(Table[2](https://arxiv.org/html/2608.00011#S4.T2)\), DLLM\-TTS attains a MOS of4\.25, behind only OpenAudio\-s1\-mini and LLASA\-3B and ahead of all other baselines\. This confirms that our objective gains translate into perceptual quality, and that block\-wise denoising boundaries do not harm naturalness or prosody\.

Data Efficiency\.Compared to autoregressive codec language models trained on 60K–250K hours, DLLM\-TTS achieves competitive intelligibility with only20K hours, a3–12×\\timesdata reduction\. We attribute this to the masked diffusion training objective \(Eq\.[5](https://arxiv.org/html/2608.00011#S3.E5)\), which exposes each sequence to diverse masking patterns across timesteps, providing implicit data augmentation compared to the single left\-to\-right ordering of autoregressive training\.

Latency\.Following the inference procedure in Section[3\.4](https://arxiv.org/html/2608.00011#S3.SS4), withT=16T=16steps per block and confidence thresholdτ=0\.6\\tau=0\.6, DLLM\-TTS achieves an RTF of0\.15\. Each block ofB=32B=32codec tokens spansB/fr=0\.64B/f\_\{r\}=0\.64s of audio, enabling low time\-to\-first\-audio and streaming synthesis after the first block is denoised\. KV caching across theKKsequential blocks further reduces latency\.

Table 2:Results on Seed\-TTS\-eval \(English\)\. MOS collected from 25 listeners under the CodecMOS\-Accent protocol\.Bold: best\.Underline: second best\.![Refer to caption](https://arxiv.org/html/2608.00011v1/x1.png)Figure 2:Visualization of the block\-wise denoising process\. \(B=32B=32,T=32T=32\)
### 4\.3Ablation Studies

Effect of Denoising Steps\.Table[3](https://arxiv.org/html/2608.00011#S4.T3)\(top\) evaluates the number of denoising stepsTTper block during inference \(Section[3\.4](https://arxiv.org/html/2608.00011#S3.SS4)\)\. Withτ=0\.6\\tau=0\.6\(Eq\.[7](https://arxiv.org/html/2608.00011#S3.E7)\), increasingTTimproves intelligibility while maintaining stable speaker similarity: WER decreases from 14\.58% to 2\.25% and CER from 7\.86% to 1\.05% when moving fromT=8T=8toT=32T=32, while SIM remains relatively stable \(0\.746–0\.765\)\. AtT=16T=16\(=B/2=B/2\), the model achieves a favorable speed–quality trade\-off with an RTF of0\.15\.

Table 3:Ablation studies\.Top: denoising stepsTT\(atB=32B=32\)\.Bottom: block sizeBB\(atT=BT=B\)\.Bold: best\.Effect of Block Size\.Table[3](https://arxiv.org/html/2608.00011#S4.T3)\(bottom\) varies block sizeBBwith matched denoising steps \(T=BT=B\)\. LargerBBraises within\-block parallelism but reduces sequential boundariesK=⌈L/B⌉K=\\lceil L/B\\rceil, weakening cross\-block conditioning under the staircase mask \(Eq\.[4](https://arxiv.org/html/2608.00011#S3.E4)\); smallerBBstrengthens autoregressive guidance but limits parallel context\. Empirically,B=32B=32is best \(WER 2\.25%, CER 1\.05%, SIM 0\.750\): smaller blocks \(B=8,16B=8,16\) raise error rates, while excessively large blocks sacrifice sequential modeling capacity and degrade intelligibility\. Speaker similarity stays stable across configurations \(0\.725–0\.750\), showing robust voice cloning\.

## 5Conclusion

We presented DLLM\-TTS, a TTS framework based on block discrete diffusion language modeling over neural audio codec tokens\. By decomposing codec token sequences into blocks and applying masked diffusion within each block while processing them sequentially, the model learns both local acoustic coherence and global text–speech alignment without requiring explicit duration modeling or phoneme\-level annotations\.

A 0\.6B\-parameter model trained on 20K hours of data achieves competitive intelligibility on the Seed\-TTS benchmark while obtaining strong speaker similarity, demonstrating improved data efficiency compared to autoregressive codec language models trained on 60K–250K hours\. The block\-parallel inference strategy achieves a real\-time factor \(RTF\) of 0\.15, enabling practical real\-time speech synthesis\.

These results suggest that block discrete diffusion language models offer a scalable alternative to autoregressive speech models, combining the parallel generation advantages of diffusion with the sequential structure needed for stable text–speech alignment\.

## 6Use of Generative AI Disclosure

In preparing this manuscript, the authors used generative AI tools for language refinement \(rephrasing and improving the clarity of author\-written text\) and as a coding assistant \(helping write and debug software for experiments and analysis\)\. All research contributions, including the methodology, experimental design, results, and scientific claims, are the authors' own\. The authors reviewed and verified all AI\-assisted text and code, and take full responsibility for the content of this paper\.

## References