Paper deep dive
Training a Compact, Strong, and Realistic Text-to-Speech Model
Runyuan Cai, Yu Lin, Yiming Wang, Xiaodong Zeng
Intelligence
Status: not_run | Model: - | Prompt: - | Confidence: 0%
Entities (0)
Relation Signals (0)
No relation signals yet.
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Text-to-speech (TTS) systems are conventionally optimized for intelligibility and speaker similarity. However, neither objective captures the prosodic detail that distinguishes human performances from synthetic speech, leaving a realism gap that larger models and higher-quality data can narrow but not close. In the absence of an objective that directly targets this gap, realism remains a persistent challenge in TTS. We present Audio8-TTS, a compact 0.6B-parameter model for zero-shot voice cloning that achieves state-of-the-art performance in intelligibility and speaker similarity, with particularly strong results in perceptual naturalness. To align the training signal with human perception rather than signal-level fidelity, we construct a quality ladder from widely used open-source TTS systems and train a realism critic to assess generated speech against matched human recordings of the same sentences. An identity-erasure objective discourages the critic from relying on generator fingerprints. We train the model in three stages: multilingual supervised fine-tuning; group relative policy optimization (GRPO) with word error rate (WER) and speaker similarity (SIM) rewards; and GRPO guided by the realism critic. The final stage improves perceived prosodic naturalness and speaker similarity while preserving WER performance.
Tags
Links
- Source: https://arxiv.org/abs/2610.14776v1
- Canonical: https://arxiv.org/abs/2610.14776v1
Trouble viewing inline? Open PDF directly ā
Full Text
49,051 characters extracted from source content.
Expand or collapse full text
Training a Compact, Strong, and Realistic Text-to-Speech Model Runyuan Cai Yu Lin Yiming Wang Xiaodong Zeng Affiliation: AutoArk October 10, 2026 Abstract Text-to-speech (TTS) systems are conventionally optimized for intelligibility and speaker similarity. However, neither objective captures the prosodic detail that distinguishes human performances from synthetic speech, leaving a realism gap that larger models and higher-quality data can narrow but not close. In the absence of an objective that directly targets this gap, realism remains a persistent challenge in TTS. We present Audio8-TTS, a compact 0.6B-parameter model for zero-shot voice cloning that achieves state-of-the-art performance in intelligibility and speaker similarity, with particularly strong results in perceptual naturalness. To align the training signal with human perception rather than signal-level fidelity, we construct a quality ladder from widely used open-source TTS systems and train a realism critic to assess generated speech against matched human recordings of the same sentences. An identity-erasure objective discourages the critic from relying on generator fingerprints. We train the model in three stages: multilingual supervised fine-tuning; group relative policy optimization (GRPO) with word error rate (WER) and speaker similarity (SIM) rewards; and GRPO guided by the realism critic. The final stage improves perceived prosodic naturalness and speaker similarity while preserving WER performance. 1 Introduction Autoregressive language models over discrete audio tokens have become a leading paradigm for speech generation. Neural audio codecs encode waveforms as compact sequences of discrete codes [36, 8], enabling language models trained on these codes to synthesize speech at a quality once attainable only with dedicated acoustic pipelines [1]. This representation also supports unified modeling of recognition, synthesis, and conversion within a single model [5], and subsequent work has extended this line of research to decoupled continuous representations [31]. These advances show that compact speech models can perform demanding audio tasks when trained with suitable objectives. Accordingly, directly optimizing for perceptual realism applies this principle to a quality axis absent from conventional fidelity objectives, enabling compact models to better exploit their existing capacity. How to design a realism oriented training phase remains a key challenge. TTS systems are commonly evaluated and optimized for two properties: lexical intelligibility, measured by word error rate (WER), and speaker similarity (SIM), which measures how closely the synthesized voice matches a reference. These metrics may appear to cover the main requirements of TTS, but each is a narrow proxy. WER measures whether the intended words are recognized, not how naturally they are delivered. SIM measures speaker identity in a compact representation, leaving much of the utteranceās expressive realization unconstrained. Prosodic details such as delivery style, rhythm, and timing therefore remain largely outside the score. A model can achieve low WER and high SIM while still sounding synthetic. Optimizing only these metrics is like judging athletic ability by bodybuilding standards: it rewards muscularity and definition, but not necessarily athletic performance. More data and greater model capacity may narrow this realism gap, but neither directly optimizes perceptual naturalness. Reinforcement learning, often used to enhance model expressiveness, inherits the same limitation when guided only by WER and SIM: it can improve measured fidelity without necessarily making speech sound more natural. To address this gap, we present Audio8-TTS, a compact 0.6B-parameter model for zero-shot voice cloning. Its decoder follows the fastāslow dual-autoregressive formulation used in recent codec language models [32, 20]. The slow branch uses Qwen2.5 0.5B as its backbone, with its vocabulary extended to include audio tokens. A shallow fast branch predicts the remaining codebooks. Together with our training framework, this architecture enables Audio8-TTS to achieve state-of-the-art performance in intelligibility, speaker similarity, and perceptual naturalness, despite having far fewer parameters than substantially larger models. Training proceeds in three stages: supervised fine-tuning (SFT) on multilingual speech; group relative policy optimization (GRPO) [30] with WER and SIM rewards to stabilize content and speaker identity; and a final GRPO stage guided by extra realism critic to enhance perceptual naturalness. We train the critic on a quality ladder of outputs from widely used open-source TTS systems, ranked through a private blind listening study, with matched human recordings as the top rung. Unlike public arenas, where each system speaks in its own voice, our study holds speaker identity and sentence content constant: every system clones the same reference voice and reads the same target sentence. The critic scores each TTS candidate against the matched human recording of that sentence. System-specific fingerprints can reveal which generator produced a sample without revealing how natural it sounds. We therefore introduce an identity-erasure strategy to keep the criticās ratings focused on prosodic realism. Another challenge is constraining the output duration within a plausible range. Since progressive pause compression is part of our ladder construction strategy, candidates with less pronounced pauses is designed to receive lower realism scores. So optimization can favor longer outputs regardless of the naturalness of speaking rate, providing another potential shorcut. Since the WER and SIM are largely insensitive to duration, an explicit duration constraint is further introduced as the solution. In this technical report, we introduce a training framework for compact zero-shot voice cloning. The framework uses a learned realism critic as a reward model, complementing WER and SIM rewards with a signal for perceptual naturalness. Applied to Audio8-TTS, a 0.6B-parameter model, this framework achieves state-of-the-art results in intelligibility, speaker similarity, and perceptual naturalness. 2 Related Work 2.1 Neural Text-to-Speech and Discrete Audio Tokens Early end-to-end neural text-to-speech systems mapped text to intermediate acoustic features [34]; later systems generated waveforms directly [16]. Neural audio codecs introduced another way to represent speech, encoding waveforms as short sequences of discrete codes [36, 8]. Self-supervised speech models also provide discrete units for autoregressive modeling [14, 17]. Representing speech as a token sequence makes synthesis a language-modeling task, as in recent zero-shot systems [1]. The same approach also allows a single autoregressive model to handle recognition, synthesis, and conversion through instruction-driven task induction. We adopt this discrete-token approach and focus on a compact model for synthesis. Its main design challenges are efficiently decoding codebook residuals and maintaining perceptual audio quality. 2.2 Preference Optimization for Generative Models Reinforcement learning from human feedback uses preference signals to align generative models with human judgments [7, 26]. Proximal policy optimization (PPO) [29] is widely used for this purpose, but requires an auxiliary value model. Group relative policy optimization (GRPO) avoids that requirement by normalizing rewards across a group of responses to the same prompt [30]; it is now commonly used to post-train language models. Applying these methods to speech adds a practical challenge: each rollout must first be synthesized and then scored. The reward must also reliably distinguish samples that differ mainly in prosody. 2.3 Reward Models and Listening Tests for Speech Objective quality predictors estimate an absolute mean opinion score (MOS) from a single utterance [11]. They can rank degraded samples, but their training domains are narrow, and they do not directly compare a candidate with a human performance. Listening arenas compare systems head-to-head, but each system speaks in its own voice, even when the text is held fixed. Their rankings can therefore reflect speaker preference as well as synthesis quality. Our reward model instead scores each candidate against a matched human recording of the same sentence, holding content and speaker identity constant. It combines a frozen self-supervised frontend [6] with explicit prosodic features, rather than predicting an absolute quality score. 3 A Compact Dual-Autoregressive Model Our decoder adopts the fastāslow dual-autoregressive design used in recent codec language models [32, 20]. Figure Figure 1 shows the full pipeline. The code, model weights, and implementation and deployment details are released in the Audio8-TTS repository [2]. We describe the two branches and their computational cost below. Figure 1: Overview of Audio8-TTS. The slow autoregressive branch predicts text tokens and the first audio codebook at the frame rate, while a shallow fast branch predicts the remaining nine codebooks in parallel. The audio decoder remains frozen. 3.1 Discrete Audio Tokens A fixed audio tokenizer maps each waveform to T frames at 21.5 Hz21.5\,Hz. Each frame comprises ten discrete codes, one drawn from each of ten codebooks with size 40964096. Speech synthesis is therefore reduced to generating a 10ĆT10Ć T matrix of discrete tokens, after which a frozen audio decoder reconstructs the waveform. Both the tokenizer and the decoder remain fixed throughout language-model training. 3.2 Slow and Fast Branches Generation is factorized along the frame and codebook axes. The slow branch operates autoregressively at the frame rate, predicting the text tokens together with the first audio codebook. It is based on a Qwen2.5 0.5B language model [27] augmented with audio-token embeddings, with a hidden size of 896 and 24 layers. Conditioned on the hidden state produced by this branch, a shallow fast branch with four layers predicts the remaining nine codebooks of each frame in parallel. Because the slow branch operates only at 21.5 Hz21.5\,Hz, the autoregressive sequence remains short, while codebook-level detail is handled by a branch roughly an order of magnitude shallower. 3.3 Parameter Budget and Inference The autoregressive decoder has 601M parameters, excluding the frozen codec. The slow branch has 24 layers, and the fast branch has 4 layers; both have a width of 896. They use grouped-query attention with 14 query heads and two keyāvalue heads. The decoder also includes ten codebook embedding tables, each with 4096 entries. At inference, each frame passes through the slow branch once and the fast branch once. The fast branch predicts nine codebooks in parallel, adding no autoregressive steps. Thus, sequential computation occurs only at the frame level. For CPU inference, the released ONNX implementation uses the same schedule and applies weight-only INT4 quantization to both autoregressive branches. Activations, the KV cache, and the codec remain in FP16. In the reported Apple M2 configuration, memory use is approximately 1.0 GiB after loading and peaks at 1.1ā1.2 GiB. The online model files total about 572 MiB. 4 Training for Realism Training has three stages: multilingual supervised fine-tuning, GRPO with WER and SIM rewards, and a final GRPO stage guided by the realism critic as well as auxiliary WER and SIM reward components. 4.1 Supervised Fine-tuning For supervised fine-tuning, each example pairs reference and target recordings from the same speaker and includes a transcript for each. The training set combines cleaned Emilia [13] and YODAS2 [18] with a private collection covering speakers and recording conditions underrepresented in public datasets. We exclude low-quality recordings, compare transcripts against automatic speech recognition (ASR) output, and filter samples by language, duration, and ASR error rate. The target waveform is encoded in the discrete token space defined in Sec. 3; the reference audio and transcript use the inference-time conditioning format. We train both the two branches with cross-entropy: the slow branch predicts text and the first codebook at the frame rate, while the fast branch predicts the remaining 9 codebooks in parallel. Training begins with the slow branch alone, proceeds to the fast branch with the slow branch frozen, and ends with joint fine-tuning. This SFT stage yields intelligible, speaker-consistent speech across the languages in the training corpus. The preference-optimization stages then preserve and enhance these strengths, while addressing the perceptual unnaturalness that SFT leaves unresolved (see Sec. 1). 4.2 Learning a Reference-Based Realism Critic After SFT, the first GRPO stage optimizes WER and SIM; a second adds a reference-based reward for perceptual realism. We first describe the method for constructing the critic that provides this reward. Quality-tier construction. A naive human-versus-synthetic classifier cannot rank synthetic speech by quality. When its synthetic examples come from a single generator, source identity is confounded with the class label: the classifier can exploit generator-specific cues rather than learn to assess speech quality. To provide graded supervision, we construct a quality ladder from outputs of six open-source systems: Fish-Speech 1.5 [20], IndexTTS-2 [38], Qwen3-TTS [15], Higgs Audio v3 [4], Fish Audio S2 [21], and Spark-TTS [33]. We rank the six TTS systems in a private blind listening study. In each pairwise comparison, listeners hear outputs for the same sentence and reference voice, without being told which systems produced them. Holding content and speaker identity fixed is essential: otherwise, as in public arenas where systems speak in their own voices, speaker preference can influence the quality ranking. We aggregate the pairwise judgments into an Elo-style ranking and place the matched human recording at the top. This ordered set forms the quality ladder used in our experiments. All six systems contribute to the listening study and appear on the resulting quality ladder, but only four provide critic-training data. We reserve Fish Audio S2[21] and Spark-TTS[33] for evaluation, testing whether the critic can score unseen systems in a way that preserves their listening-study ranking and places their scores within the ranges expected for their respective ladder tiers. Critic design. The critic scores each synthesized utterance against the matched human recording of the same sentence. It is trained to rank utterances within sentence-specific groups, with the human recording anchoring the top of the quality ladder; its scores do not represent absolute quality. We pass both the synthesized utterance and its human reference through a frozen WavLM-large encoder [6], extract layer 8, and pool the frames to 25 Hz. We also provide explicit pitch, energy, and voicing trajectories for both. After monotonic-band alignment, multiscale residual features are mapped to a learned distance d. The critic returns the bounded score s=āBātanhā”(d/B),B=30,s=-B (d/B), B=30, (1) where sā(ā30,0]sā(-30,0] and larger values indicate speech closer to the human reference. The score is not calibrated as a mean opinion score (MOS). Since policy optimization normalizes rewards within each group, only relative scores within a group affect the update. Only the scoring head is trained. It has roughly 3M parameters; the 316M-parameter WavLM encoder remains frozen. This keeps scoring inexpensive relative to generating the rollout it evaluates. We use the same feature-extraction pipeline during training and inference, including the WavLM layer, temporal pooling, prosodic features, and waveform preprocessing. Because the criticās scores depend on this representation, we store a hash of the pipeline configuration with the critic and verify it at load time. Scoring is disabled if the active configuration differs. Critic training. Each training group is indexed by a sentence and a reference voice. It contains the matched human recording y and, for each of the K systems used to train the critic, synthesized utterances xi,ā=0L\x_i, \_ =0^L of the same sentence in the same voice, at degradation levels ā=0,ā¦,L =0,ā¦,L. The reference voice conditions each system and fixes speaker identity across systems; the human recording serves as the perceptual anchor for scoring. For brevity, we write sā”(xi,ā)s(x_i, ) for the reference-conditioned score sā”(xi,ā,y)s(x_i, ,y). Constructing the degradation levels. We supplement the system-quality ordering with a second comparison axis: progressively weakened internal rhythm. Here, rhythm refers to the timing and scale of auditory pauses. We detect them, scale their durations by a factor, and resynthesize each utterance using Praatās duration manipulation, based on pitch-synchronous overlap-add [23]. Other features including pitch, spectral envelope, and lexical content are preserved during this operation. Let djd_j denote the original duration of detected pause j. At level ā , its duration is dj,ā=aāādj,1=a0ā„a1ā„āÆā„aL=0.d_j, =a_ d_j, 1=a_0ā„ a_1ā„Ā·sā„ a_L=0. Thus, the pause pattern is unchanged at level 00 and all detected pauses are removed at level L, forming a chain of rhythm degradation. To check that the effect is not specific to Praat, we repeat the construction with WORLD-based resynthesis [22]. Same score trend indicates that the criticās response is not specific to Praatās resynthesis procedure. Monotonicity within a system. As pauses are progressively compressed, the score for each system should decrease by at least a margin m: āmono=āi=1Kāā=0Lā1maxā”(0,sā”(xi,ā+1)āsā”(xi,ā)+m).L_mono= _i=1^K _ =0^L-1 \! (0,\,s(x_i, +1)-s(x_i, )+m ). (2) The hinge loss vanishes once the score drop reaches m. This margin sets a minimum score gap between adjacent degradation levels, giving pause-related changes a local scale that the cross-system ranking loss does not provide. Ranking across systems. We denote by P the set of system pairs whose relative quality is reliably established by the listening study. For each pair in this set, we train the critic to preserve the observed ordering at every degradation level: ārank=āā=0Lā(i,j)āBCEā”(Ļā”(sā”(xj,ā)āsā”(xi,ā)T),piāj,ā),L_rank= _ =0^L _(i,j) BCE\! (Ļ\! ( s(x_j, )-s(x_i, )T ),\,p_ij, ), (3) where piāj,āp_ij, is a soft target derived from the listening-study quality tiers. As rhythm flattening increases, the target moves toward 1/21/2, allowing the score gap between systems to narrow without reversing their order. The ranking loss preserves the listening-study order; the monotonicity loss separately requires scores to fall as each systemās rhythm is flattened. Erasure of identity. At level ā=L =L, all detected pauses have been removed. We treat any remaining score differences across systems as possible generator-identity leakage and penalize them directly: āera=āi<j|sā”(xi,L)āsā”(xj,L)|.L_era= _i<j |s(x_i,L)-s(x_j,L) |. (4) The ranking loss may still preserve some score separation at this endpoint, as specified by piāj,Lp_ij,L, while āeraL_era pulls the scores together. At level ā=0 =0, the ranking target preserves the listening-study order. In practice, the score spread among fully flattened variants is a small fraction of that among the original outputs. The full objective is ā=ārank+Ī»māāmono+Ī»eāāera+ānuis,L=L_rank+ _mL_mono+ _eL_era+L_nuis, where ānuisL_nuis penalizes scores predictable from duration or energy. Validation. We evaluate the critic on 5,600 held-out sentenceāvoice groups. Each group contains one utterance from each of the four training systems at each of four degradation levels, for a total of 89,600 synthesized utterances, in addition to the matched human references. Systems used for training. On these held-out groups, the score distributions follow the quality tiers established by the listening study: the human references occupy the top of the range, followed by the four systems in ladder order. The critic thus recovers the listening-study ranking on sentenceāvoice groups excluded from training. Systems excluded from training. Fish Audio S2 and Spark-TTS are not used to train the critic. Their score distributions nevertheless align with the tiers assigned to them in the listening study, rather than collapsing into the tiers of the training systems. The critic therefore preserves the studyās quality distinctions for generators it has not seen during training. The scoring path is deterministic and performs no resynthesis, so repeated scoring of the same utterance returns the same value. This consistency is required during policy optimization, which scores every rollout. 4.3 Optimizing the Policy with GRPO The final two training stages use group relative policy optimization (GRPO) [30]. For each prompt, we sample a group of N rollouts from the current policy, decode them to audio, and score them. The stage-specific reward axes are normalized and combined into a scalar qiq_i, as described below. GRPO then centers and scales this scalar within the group: Ai=qiāqĀÆĻq+Ļµ,qĀÆ=1Nāāj=1Nqj,Ļµ=10ā6,A_i= q_i- q _q+Īµ, q= 1N _j=1^Nq_j, Īµ=10^-6, (5) where Ļq _q is the standard deviation of the combined scores. The additive ĻµĪµ stabilizes the normalization when the scores have little variation. We then subtract the duration cost to obtain the coefficient used in the policy update: A~i=AiāĪ»āCi, A_i=A_i-Ī» C_i, (6) where CiC_i is the normalized duration cost defined in Sec. 4.4. The policy update uses A~i A_i to weight the log-probabilities of rollout i. We recompute these log-probabilities with a forward pass over the sampled sequence rather than reading cached generation-time values. The scoring models remain frozen, and no gradient is propagated through them. During training, scoring runs in a separate, persistent service that transcribes and embeds decoded audio while the policy samples and updates. Fidelity stage. The first GRPO stage optimizes lexical accuracy and speaker similarity. We compute WER with Whisper-large-v3 [28] for non-Chinese input and an in-house recognizer for Chinese [10]. We map WER to reward with a clipped linear function, so lower WER yields higher reward: riASR=clipā”(1āWERi2.0, 0, 1),r_i^ASR=clip\! (1- WER_i2.0,\,0,\,1 ), (7) where WER is represented as a decimal fraction (e.g., 2%=0.022\%=0.02). We compute SIM as the cosine similarity between candidate and reference embeddings from a WavLM-based speaker encoder [6], then rescale it to [0,1][0,1]: riSIM=clipā”(SIMi0.85, 0, 1).r_i^SIM=clip\! ( SIM_i0.85,\,0,\,1 ). (8) The fidelity score combines the two rewards before group-level standardization: sifid=0.7āriASR+0.3āriSIM.s_i^fid=0.7\,r_i^ASR+0.3\,r_i^SIM. (9) Realism stage. The final GRPO stage combines the fidelity reward with the realism criticās score. For rollout i in group g, let sifids_i^fid denote its fidelity score and sipross_i^pros its critic score. We center each score by its mean across the group and scale it by the larger of the group standard deviation and an axis-specific minimum: zifid z_i^fid =sifidāsĀÆgfidmaxā”(Ļgfid, 0.005), = s_i^fid- s_g^fid ( _g^fid,\,0.005), (10) zipros z_i^pros =siprosāsĀÆgprosmaxā”(Ļgpros, 0.3), = s_i^pros- s_g^pros ( _g^pros,\,0.3), qi q_i =zifid+wzipros,w=1, =z_i^fid+w\,z_i^pros, w=1, where sĀÆg s_g and Ļg _g are the mean and standard deviation within group g. The floors prevent small group variances from amplifying either score. The combined scalar qiq_i is the reward passed to the GRPO normalization. With w=1w=1, the two axes have equal coefficients, although the floors can make their effective magnitudes differ. We combine the axes before computing the GRPO advantage. Computing a separate advantage for each would normalize each axis independently, potentially giving a low-variance signal disproportionate influence. Standardizing the critic score also prevents its prompt-dependent raw spread from determining its contribution to the combined reward. 4.4 Bounding Temporal-Scale Drift The pause-compression ladder metioned in Sec. 4.2 provides one controlled cue in critic training: within each system, progressively compressed pauses receive lower scores. This teaches the critic to treat pause structure as one aspect of prosodic realism but can also encourage a favor for longer total duration, since preserving pauses increase the length of utterance. Hence we constrain temporal scale separately with a one-sided log penalty beyond a 35% margin over the matched reference. Let Ļi _i be the ratio of rollout duration to reference duration, with threshold k=1.35k=1.35. The one-sided log penalty is ui=clipā”(logā”Ļik, 0,c),Ci=uiāuĀÆgmaxā”(Ļgu, 0.1),u_i=clip\! ( _ik,\,0,\,c ), C_i= u_i- u_g ( _g^u,\,0.1), (11) where c caps the penalty, and uĀÆg u_g and Ļgu _g^u are the mean and standard deviation of uiu_i within rollout group g. Thus, ui=0u_i=0 below the duration threshold and increases logarithmically above it. If all rollouts in a group satisfy the threshold, the centered cost is zero for the entire group. Dual ascent adjusts Ī» toward a 15% rate of rollouts exceeding the threshold. The cost enters the policy update as defined in Sec. 4.3. On a fixed evaluation set, duration stays within 6% of its initial value, with fewer than 5% of utterances exceeding the reference by more than 50%. Critic scores and speaker similarity nevertheless improve, supporting gains in delivery rather than duration. 5 Analysis We report results on the nine languages in the CV3 benchmark [9]āChinese, English, Japanese, Korean, German, French, Spanish, Italian, and Russianāand on thirteen additional languages: Arabic, Finnish, Hindi, Indonesian, Malay, Dutch, Norwegian, Polish, Portuguese, Swedish, Thai, Turkish, and Vietnamese. Per-language results are given in Appendix A. 5.1 Data Mixture Chinese and English each make up about 30% of the supervised fine-tuning data; each other language contributes at most 3%. In the GRPO stage optimized with WER and speaker-similarity rewards, about two-thirds of the prompts are in the nine CV3 languages, with roughly 7% from each. The remaining prompts are in thirteen additional languages, with roughly 3% from each language except Thai, which accounts for about 1%. For the realism stage, the prompts corpus uses Chinese and English in roughly equal proportions. 5.2 Stage-wise Results On the nine CV3 languages, mean error rate falls from 5.41% after SFT to 5.01% after WER/SIM optimization and 4.86% after realism optimization. The WER/SIM stage lowers error in every language. Speaker similarity also improves across stages, with most of the increase occurring during realism optimization. On the thirteen additional languages, WER/SIM optimization yields no aggregate improvement: error rate rises slightly from 22.36% to 22.53%, while SIM changes from 0.638 to 0.637. The aggregate changes are small because improvements in some languages are offset by regressions in others. By contrast, realism optimization improves both aggregate metrics, with lower error rates and higher SIM in all thirteen languages. SFT GRPO (WER/SIM) GRPO (realism) CV3, error rate (%) ā 5.41 5.01 4.86 CV3, SIM ā 0.639 0.643 0.661 Add. languagesā , error rate (%) ā 22.36 22.53 20.77 Add. languagesā , SIM ā 0.638 0.637 0.679 Table 1: Stage-wise results on the nine CV3 languages and thirteen additional languages. Error rate is WER, except for Chinese, Japanese, and Korean, where it is CER. The same recognizer and speaker encoder are used throughout. ā The additional 13 languages are outside CV3 and less represented in the training data; their results are reported separately. Params. Seed-TTS (en/zh) CV3 (9) MiniMax (24) Model (B) Error ā SIM ā Error ā SIM ā Error ā SIM ā Seed-TTS [1] ā¼ 1.68a1.68^a 0.779a0.779^a ā¼ ā¼ ā¼ ā¼ MiniMax [37] ā¼ 1.24a1.24^a 0.738a0.738^a ā¼ ā¼ 3.77a3.77^a 0.766a0.766^a IndexTTS-2.5 [19] 0.81p,ā0.81^p, 1.66a1.66^a 0.852a,f0.852^a,f ā¼ ā¼ ā¼ ā¼ Qwen3-TTS-12Hz-0.6B [15] 0.6 1.12a1.12^a ā¼ ā¼ ā¼ ā¼ ā¼ Qwen3-TTS-1.7B [15] 1.7 1.01a1.01^a ā¼ 7.73h7.73^h ā¼ 27.41h27.41^h ā¼ Higgs Audio v3 [4] 4.65p4.65^p 1.11h1.11^h ā¼ 4.41h4.41^h ā¼ 2.74h2.74^h ā¼ Higgs Audio v2 [3] 4.7p4.7^p 1.17e1.17^e 0.693e0.693^e 4.49e4.49^e ā¼ ā¼ ā¼ MOSS-TTS-v1.5 [25, 12] 8.5p8.5^p 1.52c1.52^c 0.761c0.761^c 6.11h6.11^h ā¼ 3.78h3.78^h ā¼ FireRedTTS-2 [35] 1.7 1.72h1.72^h ā¼ 19.20h19.20^h ā¼ 12.52h12.52^h ā¼ Fish S2 Pro [21] 4.6p4.6^p 1.32e1.32^e 0.692e0.692^e 4.49e4.49^e ā¼ 4.40a4.40^a 0.780a0.780^a Spark-TTS [33] 0.5 1.59a1.59^a 0.628a0.628^a ā¼ ā¼ ā¼ ā¼ CosyVoice3-1.5B [9] 1.5 1.67c1.67^c 0.750c0.750^c 6.89c6.89^c ā¼ ā¼ ā¼ VoxCPM2 [39] 2.3p2.3^p 1.41c1.41^c 0.774c0.774^c 5.35c5.35^c ā¼ ā¼ ā¼ Audio8-TTS (ours) 0.6 1.25a1.25^a 0.704a0.704^a 4.86 0.661 5.89 0.754 Table 2: Results on Seed-TTS, CV3, and MiniMax. Error rates (%) are macro-averaged over languages. Seed-TTS averages English WER and Chinese CER, excluding zh-hard. ā¼ : unavailable. aMeans computed from reported per-language results. cResults compiled in the VoxCPM repository [24]. eOur evaluation. hBoson AIās evaluation [4]. fFunASR-based SIM. pParameter counts computed from checkpoint tensors, excluding codec weights and buffers; otherwise, counts are taken from model releases. āExcludes the separate acoustic decoder. Parameter-counting conventions differ across models. In addition to the nine-language CV3 benchmark, we evaluate the final model on Seed-TTS [1] and the 24-language MiniMax test set [37]. Table 2 compares our results with other TTS systems. We reevaluate Fish S2 Pro on Seed-TTS and CV3 because its reported results use a different text-normalization procedure. We also evaluate Higgs Audio v2, for which the required results were not available. For the remaining baseline results, we use the original reports, Boson AIās evaluation [4], and results compiled in the VoxCPM repository [24]. Evaluation protocols differ across sources; in particular, SIM scores computed with different speaker encoders are not directly comparable. Full-suite results may include languages outside a modelās supported set. For Qwen3-TTS, the Seed-TTS results use the 12Hz Base models. Bosonās CV3 and MiniMax results do not specify the tokenizer variant. Per-language CV3 error rates for these systems, including the two hard subsets, are reported in Table 6. Across the 22 CV3 languages, the final model has higher speaker similarity than the SFT model in every language. Recognition error is lower in all but Korean and Thai; for both, it remains within one percentage point of the post-SFT result (Table 3). On MiniMax, Cantonese is the only language with an error rate above 20% (Table 4). This likely reflects the training dataās strong emphasis on Mandarin Chinese, as well as the substantial overlap between the token spaces for Cantonese and Mandarin. 5.3 Blind Human Evaluation We conducted a blind listening study to compare outputs from the SFT, WER/SIM, and realism stages. The realism-stage model received a substantially higher Elo rating than either of the other two models, while the SFT and WER/SIM models received similar ratings. 6 Limitations and Conclusion Limitations. The realism critic requires a matched human reference for each sentence. Its scores are relative: they rank candidates for the same prompt according to the ordering established by the listening study, but do not provide an absolute measure of naturalness or replace a content check. Though the critic ranks overall synthesis quality across systems, its controlled signal for separating prosodic cues from generator identity comes from pause-flattened variants. These variants isolate pause timing; they do not provide equally targeted supervision for pitch, energy, or other prosodic cues. The duration penalty reduces length drift but does not eliminate it. Preference optimization has a different limit: it can make existing prosodic patterns more or less likely, but cannot teach the model patterns it never produces. The 13 additional languages account for a smaller share of the training data, and their recognition error remains substantially higher than that on the nine CV3 languages. We do not evaluate audio artifacts such as buzzing or vocoder noise and the metrics used in evaluation would not capture them. Conclusion. We presented Audio8-TTS, a 0.6B dual-autoregressive model trained in three stages: supervised fine-tuning, GRPO with word error rate and speaker similarity rewards, and GRPO with a reference-anchored realism critic. The critic is trained against a quality ordering established through a blind listening study, with an identity-erasure objective to keep its score from tracking generator identity. On the nine CV3 languages, error rate falls at each stage and mean speaker similarity rises. On the 13 additional languages, the WER/SIM stage leaves aggregate results nearly unchanged, while the realism stage improves both metrics in every language, despite using only Chinese and English prompts. In a separate blind listening study, the realism-stage model receives a substantially higher Elo rating than the SFT and WER/SIM models, whose ratings are similar. The duration constraint also keeps output duration close to its initial level as critic scores and speaker similarity improve. Acknowledgments We thank the developers of Fish-Speech 1.5 [20], IndexTTS-2 [38], Qwen3-TTS [15], Higgs Audio v3 [4], Fish Audio S2 [21], and Spark-TTS [33] for releasing the systems used in our quality ladder. The Fish Audio S2 implementations also underpin much of our training and inference code. We thank the CosyVoice 3 [9] and Seed-TTS [1] teams for their evaluation suites, and the authors of Whisper [28] and WavLM [6] for releasing the models used for scoring. References [1] P. Anastassiou, J. Chen, J. Chen, Y. Chen, Z. Chen, Z. Chen, et al. (2024) Seed-TTS: a family of high-quality versatile speech generation models. In arXiv preprint arXiv:2406.02430, Cited by: Appendix A, Ā§1, Ā§2.1, Ā§5.2, Table 2, Acknowledgments. [2] Audio8-TTS Team (2026) Audio8-TTS: code, model weights, and implementation details. Note: https://github.com/Edge0-AI/Audio8_TTS Cited by: Ā§3. [3] Boson AI (2025) Higgs audio V2. Note: https://huggingface.co/bosonai/higgs-audio-v2-generation-3B-base Cited by: Appendix A, Appendix A, Table 2. [4] Boson AI (2026) Higgs audio v3 TTS: conversational speech for voice AI from Boson AI. Note: https://huggingface.co/bosonai/higgs-audio-v3-tts-4b Cited by: Ā§4.2, Ā§5.2, Table 2, Table 2, Table 2, Acknowledgments. [5] R. Cai, Y. Lin, Y. Wang, C. Fu, and X. Zeng (2026) Unifying speech recognition, synthesis and conversion with autoregressive transformers. arXiv preprint arXiv:2601.10770. Cited by: Ā§1. [6] S. Chen, C. Wang, Z. Chen, Y. Wu, S. Liu, Z. Chen, J. Li, N. Kanda, T. Yoshioka, X. Xiao, et al. (2022) WavLM: large-scale self-supervised pre-training for full stack speech processing. IEEE Journal of Selected Topics in Signal Processing. Cited by: Ā§2.3, Ā§4.2, Ā§4.3, Acknowledgments. [7] P. F. Christiano, J. Leike, T. B. Brown, M. Martic, S. Legg, and D. Amodei (2017) Deep reinforcement learning from human preferences. In Advances in Neural Information Processing Systems (NeurIPS), Cited by: Ā§2.2. [8] A. DĆ©fossez, J. Copet, G. Synnaeve, and Y. Adi (2022) High fidelity speech compression with a single end-to-end model. arXiv preprint arXiv:2210.13438. Cited by: Ā§1, Ā§2.1. [9] Z. Du, C. Gao, Y. Wang, F. Yu, T. Zhao, H. Wang, X. Lv, H. Wang, X. Shi, K. An, et al. (2025) CosyVoice 3: towards in-the-wild speech generation via scaling-up and post-training. arXiv preprint arXiv:2505.17589. Note: Releases the CV3-Eval zero-shot benchmark Cited by: Table 5, Table 5, Appendix A, Appendix A, Table 2, Ā§5, Acknowledgments. [10] Edge0 (2026) ARK-ASR-3B. Note: https://huggingface.co/Edge0/ARK-ASR-3B Cited by: Ā§4.3. [11] C. Fu, C. Liu, D. Schnieders, Y. Peng, and Y. Wang (2019) MOSNet: deep learning based objective assessment for voice conversion. In Proc. Interspeech, Cited by: Ā§2.3. [12] Y. Gong, B. Jiang, Y. Zhao, Y. Yuan, K. Chen, Y. Jiang, C. Chang, D. Hong, M. Chen, R. Li, Y. Zhang, Y. Gao, H. Chen, K. Chen, S. Wang, X. Yang, Y. Zhang, K. Huang, Z. Lin, K. Yu, Z. Chen, J. Wang, Z. Fei, Q. Cheng, S. Li, and X. Qiu (2026) MOSS-TTS technical report. arXiv preprint arXiv:2603.18090. Cited by: Appendix A, Table 2. [13] H. He, Z. Shang, C. Wang, X. Li, Y. Gu, H. Hua, L. Liu, C. Yang, J. Li, P. Shi, Y. Wang, K. Chen, P. Zhang, and Z. Wu (2024) Emilia: an extensive, multilingual, and diverse speech dataset for large-scale speech generation. arXiv preprint arXiv:2407.05361. Cited by: Ā§4.1. [14] W. Hsu, B. Bolte, Y. H. Tsai, K. Lakhotia, R. Salakhutdinov, and A. Mohamed (2021) HuBERT: self-supervised speech representation learning by masked prediction of hidden units. IEEE/ACM Transactions on Audio, Speech, and Language Processing. Cited by: Ā§2.1. [15] H. Hu, X. Zhu, T. He, D. Guo, B. Zhang, X. Wang, Z. Guo, Z. Jiang, H. Hao, Z. Guo, X. Zhang, P. Zhang, B. Yang, J. Xu, J. Zhou, and J. Lin (2026) Qwen3-TTS technical report. arXiv preprint arXiv:2601.15621. Cited by: Appendix A, Appendix A, Ā§4.2, Table 2, Table 2, Acknowledgments. [16] J. Kim, J. Kong, and J. Son (2021) Conditional variational autoencoder with adversarial learning for end-to-end text-to-speech. In Proc. ICML, Cited by: Ā§2.1. [17] K. Lakhotia, E. Kharitonov, W. Hsu, Y. Adi, A. Polyak, B. Bolte, T. Nguyen, J. Copet, A. Baevski, A. Mohamed, and E. Dupoux (2021) Generative spoken language modeling from raw audio. Transactions of the Association for Computational Linguistics. Cited by: Ā§2.1. [18] X. Li, S. Takamichi, T. Saeki, W. Chen, S. Shiota, and S. Watanabe (2024) YODAS: youtube-oriented dataset for audio and speech. arXiv preprint arXiv:2406.00899. Note: Covers the YODAS collection and its successor release, YODAS2 Cited by: Ā§4.1. [19] Y. Li, X. Zhou, J. Wang, L. Wang, Y. Wu, S. Zhou, Y. Zhou, Y. Wang, Y. Yang, Z. Hu, S. Duan, J. Xu, J. Shu, and B. Xia (2026) IndexTTS 2.5 technical report. arXiv preprint arXiv:2601.03888. Cited by: Appendix A, Table 2. [20] S. Liao, Y. Wang, T. Li, Y. Cheng, R. Zhang, R. Zhou, and Y. Xing (2024) Fish-Speech: leveraging large language models for advanced multilingual text-to-speech synthesis. arXiv preprint arXiv:2411.01156. Note: Tech report for the Fish-Speech 1.4/1.5 model family Cited by: Ā§1, Ā§3, Ā§4.2, Acknowledgments. [21] S. Liao, Y. Wang, S. Liu, Y. Cheng, R. Zhang, T. Li, S. Li, Y. Zheng, X. Liu, Q. Wang, Z. Zhou, J. Liu, X. Chen, and D. Han (2026) Fish Audio S2 technical report. arXiv preprint arXiv:2603.08823. Cited by: Appendix A, Appendix A, Ā§4.2, Ā§4.2, Table 2, Acknowledgments. [22] M. Morise, F. Yokomori, and K. Ozawa (2016) WORLD: a vocoder-based high-quality speech synthesis system for real-time applications. IEICE Transactions on Information and Systems 99 (7), p. 1877ā1884. Cited by: Ā§4.2. [23] E. Moulines and F. Charpentier (1990) Pitch-synchronous waveform processing techniques for text-to-speech synthesis using diphones. Speech Communication 9 (5ā6), p. 453ā467. Cited by: Ā§4.2. [24] OpenBMB (2026) VoxCPM. Note: https://github.com/OpenBMB/VoxCPMCompiles published baseline results for text-to-speech systems Cited by: Table 5, Table 5, Table 6, Table 6, Ā§5.2, Table 2, Table 2. [25] OpenMOSS (2026) MOSS-TTS-v1.5. Note: https://huggingface.co/OpenMOSS-Team/MOSS-TTS-v1.5 Cited by: Appendix A, Table 2. [26] L. Ouyang, J. Wu, X. Jiang, et al. (2022) Training language models to follow instructions with human feedback. In Advances in Neural Information Processing Systems (NeurIPS), Cited by: Ā§2.2. [27] Qwen Team (2024) Qwen2.5 technical report. arXiv preprint arXiv:2412.15115. Cited by: Ā§3.2. [28] A. Radford, J. W. Kim, T. Xu, G. Brockman, C. McLeavey, and I. Sutskever (2023) Robust speech recognition via large-scale weak supervision. Proc. ICML. Cited by: Ā§4.3, Acknowledgments. [29] J. Schulman, F. Wolski, P. Dhariwal, A. Radford, and O. Klimov (2017) Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347. Cited by: Ā§2.2. [30] Z. Shao, P. Wang, Q. Zhu, et al. (2024) DeepSeekMath: pushing the limits of mathematical reasoning in open language models. arXiv preprint arXiv:2402.03300. Cited by: Ā§1, Ā§2.2, Ā§4.3. [31] F. Shen, F. Xie, J. Li, K. Xie, L. Xie, X. Tang, X. Geng, Y. Jia, Y. Hu, Y. Han, et al. (2026) FireRedAudio: a general-purpose audio language model with decoupled continuous representations for understanding and generation. arXiv preprint arXiv:2608.24168. Cited by: Ā§1. [32] C. Wang, S. Chen, Y. Wu, Z. Zhang, L. Zhou, S. Liu, Z. Chen, Y. Liu, H. Wang, J. Li, L. He, S. Zhao, and F. Wei (2023) Neural codec language models are zero-shot text to speech synthesizers. arXiv preprint arXiv:2301.02111. Cited by: Ā§1, Ā§3. [33] X. Wang, M. Jiang, Z. Ma, Z. Zhang, S. Liu, L. Li, Z. Liang, Q. Zheng, R. Wang, X. Feng, W. Bian, Z. Ye, et al. (2025) Spark-TTS: an efficient LLM-based text-to-speech model with single-stream decoupled speech tokens. arXiv preprint arXiv:2503.01710. Cited by: Appendix A, Ā§4.2, Ā§4.2, Table 2, Acknowledgments. [34] Y. Wang, R. J. Skerry-Ryan, D. Stanton, Y. Wu, R. J. Weiss, N. Jaitly, Z. Yang, Y. Xiao, Z. Chen, S. Bengio, Q. Le, Y. Agiomyrgiannakis, R. Clark, and R. A. Saurous (2017) Tacotron: towards end-to-end speech synthesis. In Proc. Interspeech, Cited by: Ā§2.1. [35] K. Xie, F. Shen, J. Li, F. Xie, X. Tang, and Y. Hu (2025) FireRedTTS-2: towards long conversational speech generation for podcast and chatbot. arXiv preprint arXiv:2509.02020. Cited by: Table 2. [36] N. Zeghidour, A. Luebs, A. Omran, J. Skoglund, and M. Tagliasacchi (2022) SoundStream: an end-to-end neural audio codec. IEEE/ACM Transactions on Audio, Speech, and Language Processing. Cited by: Ā§1, Ā§2.1. [37] B. Zhang, C. Guo, G. Yang, H. Yu, H. Zhang, H. Lei, J. Mai, J. Yan, K. Yang, et al. (2025) MiniMax-Speech: intrinsic zero-shot text-to-speech with a learnable speaker encoder. arXiv preprint arXiv:2505.07916. Note: Releases the twenty-four-language TTS-Multilingual-Test-Set Cited by: Appendix A, Ā§5.2, Table 2. [38] S. Zhou, Y. Zhou, Y. He, X. Zhou, J. Wang, W. Deng, and J. Shu (2025) IndexTTS2: a breakthrough in emotionally expressive and duration-controlled auto-regressive zero-shot text-to-speech. arXiv preprint arXiv:2506.21619. Cited by: Ā§4.2, Acknowledgments. [39] Y. Zhou, G. Zeng, X. Liu, X. Li, R. Yu, J. Gui, J. Wu, Z. Wang, X. Shen, R. Ye, Z. Zhang, J. Zhou, B. Bai, W. Sun, M. Deng, Q. Shi, Z. Wu, and Z. Liu (2026) VoxCPM2 technical report. arXiv preprint arXiv:2606.06928. Cited by: Appendix A, Appendix A, Table 2. Appendix A Per-Language Results Error rate ā Similarity ā Error rate ā Similarity ā Language SFT final SFT final Language SFT final SFT final Chinese 2.03 1.81 0.668 0.690 Hindiā 22.91 21.71 0.685 0.725 English 3.03 2.91 0.610 0.629 Indonesianā 15.56 13.74 0.629 0.668 Japanese 8.10 7.51 0.649 0.666 Malayā 19.48 17.80 0.636 0.671 Korean 4.60 4.80 0.681 0.696 Dutchā 11.78 10.87 0.652 0.692 German 3.64 3.20 0.646 0.667 Norwegianā 20.61 17.96 0.635 0.691 French 8.66 7.94 0.616 0.635 Polishā 21.68 19.23 0.635 0.674 Spanish 4.89 3.39 0.643 0.668 Portugueseā 7.89 6.87 0.640 0.677 Italian 5.54 4.93 0.619 0.650 Swedishā 20.97 19.25 0.671 0.710 Russian 8.21 7.26 0.621 0.645 Thaiā 47.97 48.57 0.626 0.669 Arabicā 23.17 21.10 0.571 0.612 Turkishā 13.38 12.57 0.612 0.650 Finnishā 33.80 32.28 0.663 0.703 Vietnameseā 31.48 28.11 0.638 0.679 Table 3: CV3 per-language results on the nine benchmark languages and the thirteen additional languages, for the model after supervision and after training. Error rates are word error rates, except for Chinese, Japanese and Korean, where they are character error rates. ā marks the languages we added, which lie outside the benchmarkās nine. Language Error rate ā Similarity ā Language Error rate ā Similarity ā Arabic 2.40 0.704 Italian 1.96 0.737 Cantonese 31.35 0.771 Japanese 5.14 0.789 Chinese 1.11 0.795 Korean 3.73 0.785 Czech 8.63 0.733 Polish 1.96 0.789 Dutch 1.72 0.705 Portuguese 2.23 0.786 English 2.16 0.774 Romanian 17.43 0.733 Finnish 5.07 0.791 Russian 3.07 0.749 French 3.20 0.666 Spanish 1.35 0.745 German 0.90 0.751 Thai 6.67 0.776 Greek 4.18 0.791 Turkish 1.56 0.765 Hindi 9.21 0.817 Ukrainian 16.34 0.694 Indonesian 2.13 0.727 Vietnamese 7.97 0.712 Table 4: MiniMax per-language results on the 24-language multilingual set, for the final model. Error rates are word error rates, and character error rates for the languages the suite scores by character: Chinese, Cantonese, Japanese, Korean, Arabic, Greek, Hindi, Thai and Vietnamese. Seed-TTS, en Seed-TTS, zh Seed-TTS, zh-hard Model WER ā SIM ā CER ā SIM ā CER ā SIM ā Seed-TTS [1] 2.25 0.762 1.12 0.796 7.59 0.776 MiniMax [37] 1.65 0.692 0.83 0.783 ā¼ ā¼ IndexTTS-2.5 [19] 1.89 0.855f0.855^f 1.43 0.848f0.848^f ā¼ ā¼ Qwen3-TTS-12Hz-0.6B [15] 1.32 ā¼ 0.92 ā¼ ā¼ ā¼ Qwen3-TTS-12Hz-1.7B [15] 1.24 ā¼ 0.77 ā¼ ā¼ ā¼ Fish S2 Pro [21] 1.607e1.607^e 0.646e0.646^e 1.038e1.038^e 0.738e0.738^e 10.149e10.149^e 0.701e0.701^e Higgs Audio v2 [3] 1.524e1.524^e 0.664e0.664^e 0.806e0.806^e 0.721e0.721^e 10.622e10.622^e 0.693e0.693^e MOSS-TTS-v1.5 [25, 12] 1.85c1.85^c 0.734c0.734^c 1.20c1.20^c 0.788c0.788^c ā¼ ā¼ CosyVoice3-1.5B [9] 2.22c2.22^c 0.720c0.720^c 1.12c1.12^c 0.781c0.781^c 5.83c5.83^c 0.758c0.758^c VoxCPM2 [39] 1.84c1.84^c 0.753c0.753^c 0.97c0.97^c 0.795c0.795^c 8.13c8.13^c 0.753c0.753^c Spark-TTS [33] 1.98 0.584 1.20 0.672 ā¼ ā¼ Audio8-TTS (ours) 1.46 0.663 1.03 0.745 12.80 0.702 Table 5: Seed-TTS results by split. WER and CER are reported as percentages. The zh-hard split is excluded from the means in Table 2. cResults compiled in the VoxCPM repository [24]. eOur evaluation. fFunASR-based SIM. SIM scores computed with different speaker encoders are not directly comparable. ā¼ : unavailable. Only systems with reported per-split results are included. The Seed-TTS modelās zh-hard results are taken from Du et al. [9]. Model Params (B) zh en hard-zh hard-en ja ko de es fr it ru Fish S2 Pro [21] 4.6 3.600e3.600^e 3.493e3.493^e 10.588e10.588^e 7.349e7.349^e 5.139e5.139^e 4.111e4.111^e 3.605e3.605^e 2.972e2.972^e 8.600e8.600^e 4.229e4.229^e 4.702e4.702^e Higgs Audio v2 [3] 4.7 3.378c3.378^c 3.404c3.404^c 10.424c10.424^c 5.754c5.754^c 4.742c4.742^c 4.260c4.260^c 3.300c3.300^c 2.929c2.929^c 9.425c9.425^c 3.555c3.555^c 5.423c5.423^c CosyVoice3-1.5B [9] 1.5 3.91c3.91^c 4.99c4.99^c 9.77c9.77^c 10.55c10.55^c 7.57c7.57^c 5.69c5.69^c 6.43c6.43^c 4.47c4.47^c 11.8c11.8^c 10.5c10.5^c 6.64c6.64^c VoxCPM2 [39] 2.3 3.65c3.65^c 5.00c5.00^c 8.55c8.55^c 8.48c8.48^c 5.96c5.96^c 5.69c5.69^c 4.77c4.77^c 3.80c3.80^c 9.85c9.85^c 4.25c4.25^c 5.21c5.21^c Audio8-TTS (ours) 0.6 1.814 2.907 10.665 4.736 7.511 4.801 3.200 3.389 7.938 4.933 7.258 Table 6: Per-language CV3 error rates (%): word error rates, except for Chinese, Japanese, Korean and the Chinese hard subset, which are character error rates. cResults compiled in the VoxCPM repository [24]. eOur evaluation. Each language contributes 500 items, except the two hard subsets (60 Chinese, 64 English). The mean over the nine benchmark languages is the CV3 (9) column in Table 2.