AutoArk's Audio8 TTS Model Outpaces Larger AI Voice Systems on Nine-Language Accuracy Test
Audio8-TTS-Realistic uses DualAR architecture and a novel realism critic to reach 4.86% word error rate
A research team at AutoArk released Audio8-TTS-Realistic today, an open-weight 0.6-billion-parameter text-to-speech model that achieves a 4.86% average word error rate across nine languages on the CV3 voice-cloning benchmark — outperforming Qwen3-TTS at 1.7 billion parameters (7.73%), CosyVoice3 at 1.5 billion parameters (6.89%), and MOSS-TTS-v1.5 at 8.5 billion parameters (6.11%). The weights and inference code are available under Apache 2.0 on HuggingFace alongside an arXiv paper describing the training methodology.
The result matters because Audio8-TTS-Realistic does not reach its benchmark position by adding parameters. Instead, it combines a dual autoregressive architecture with a three-stage reinforcement learning pipeline — including an explicitly trained realism critic built from a human listening study — a methodology not previously deployed in open-weight TTS at this scale. The model supports zero-shot voice cloning across 23 languages from a single reference audio clip with no fine-tuning required.
How the DualAR Architecture Separates Semantic and Acoustic Prediction
The model's dual autoregressive design splits synthesis into two differently sized tasks. A slow AR branch — 24 transformer layers with hidden width 896, initialized from the Qwen2.5-0.5B language model — predicts the first codebook of a 10-codebook neural audio codec at approximately 21.5 frames per second. This branch handles the semantically demanding work: rhythm, intonation, and phoneme-level articulation. Because the codec frame rate is far lower than the audio sample rate, the slow branch operates on a compact token sequence.
A fast AR branch of just four transformer layers then predicts the remaining nine codec codebooks conditioned on the slow branch's output. These codebooks carry the acoustic fine detail that determines timbre and naturalness. With semantic decisions already made by the slow branch, the fast branch can recover that acoustic detail without needing comparable capacity. The neural audio codec operates at 44.1 kHz with 4,096 entries per codebook, producing studio-resolution output. The architecture was originally introduced in Fish Audio's 4.6-billion-parameter S2 Pro model; Audio8 applies it at roughly one-seventh the parameter count.
Reinforcement Learning Adds a Realism Critic to Standard WER Training
After initial multilingual supervised fine-tuning on the Emilia dataset, YODAS2, and undisclosed private data, the team ran two GRPO reinforcement learning stages, as detailed in their arXiv preprint. The first used word error rate and speaker similarity as rewards in a 0.7-to-0.3 weighted combination, improving CV3 from 5.41% to 5.01% WER while speaker similarity rose from 0.639 to 0.643.
The second GRPO stage introduced the realism critic: a reward model trained on a private human blind listening study. The study ranked six TTS systems — Fish-Speech 1.5, IndexTTS-2, Qwen3-TTS, Higgs Audio v3, Fish Audio S2, and Spark-TTS — alongside matched human recordings using pause-compression degradation levels to construct Elo-style quality comparisons. The critic combines a WavLM-large encoder with prosodic features and includes an identity-erasure objective designed to prevent the model from learning to recognize the generator's acoustic fingerprint, which would reward mimicry over genuine naturalness. Adding the realism reward moved CV3 error to 4.86% and speaker similarity to 0.661.
Audio8 Versus 4-Billion-Parameter Open-Weight Systems on Three Benchmarks
On CV3's nine-language macro-average, Audio8 at 4.86% falls within half a percentage point of Higgs Audio v3 (4.41%) and Fish Audio S2 Pro (4.49%), both four- to 4.6-billion-parameter systems. All benchmark figures are company-reported; no independent reproduction has appeared yet. On the Seed-TTS English and Chinese benchmark, the picture reverses slightly: Audio8 scores a 1.25% error rate (a macro-average of English WER and Chinese CER) versus Qwen3-TTS-1.7B at 1.01% and Higgs Audio v3 at 1.11%, suggesting the parameter gap asserts itself in a narrower monolingual evaluation.
The MiniMax 24-language benchmark makes the limitation more explicit: Audio8 scores 5.89% overall against Higgs Audio v3 at 2.74% and Fish Audio S2 Pro at 4.40%. Broader language coverage appears to favor larger models, even when a compact system is competitive across nine carefully supported languages.
Thai, Cantonese, and the Evaluation Gaps the Paper Does Not Address
The model's Thai performance — 48.574% WER on the CV3 extended-language evaluation — is one of the highest error rates in the paper's benchmark tables. The paper attributes this to limited Thai training data. Cantonese character error rate reaches 31.349% on MiniMax, which the authors explain by the model's strong Mandarin emphasis and the substantial shared token space between Mandarin and Cantonese. Among the nine primary CV3 languages, French (7.938% WER) and Japanese (7.511% WER) are the weakest performers.
The paper publishes no Mean Opinion Score or human naturalness evaluation on the final model, making it impossible to independently assess whether the realism critic produced perceptible improvements beyond the WER gains. The model card also contains no discussion of voice-cloning misuse risks or content policies — a significant omission given that a compact Apache 2.0 voice-cloning model is immediately deployable in adversarial contexts including voice fraud and audio impersonation.
For production deployment, the release includes an SGLang-Omni adapter providing paged attention, dynamic batching, and an OpenAI-compatible /v1/audio/speech endpoint, allowing integration into existing TTS infrastructure without modification.
The Realism Critic Framework as the Article's Lasting Contribution
The more durable contribution from Audio8-TTS-Realistic may be the realism critic framework itself rather than any single benchmark position. If a quality-ladder-based reward model can meaningfully improve perceptual naturalness at 0.6B scale — independent of additional pretraining data — that approach is available to any team building TTS post-training pipelines, not just those with 4-billion-parameter budgets. The Apache 2.0 license places the weights and methodology within reach of independent researchers who can test that claim, and their results on naturalness evaluation will determine how much of Audio8's benchmark advantage transfers to real-world deployment.