Fish Audio Drama 3 Tops Pronunciation Evaluation, Open-Weight TTS Prices Continue to Fall
In a new Artificial Analysis evaluation, Fish Audio Drama 3 ranks first in pronunciation robustness and is the top open-weight model in both speech arenas. At $15 per million characters, it is cheaper than every model ranked above it on the Provider Voice leaderboard. A strong overall score does not mean it has mastered Chinese polyphonic characters, numbers, and mixed Chinese-English reading.
TTS leaderboards often focus on whether the voice sounds human. Once you move into production, the first problems are often pronunciation. If the stress on one word is wrong, the entire audio clip has to be re-recorded.
In this round of evaluation, Artificial Analysis put Fish Audio's new model Drama 3 through two Speech Arenas and the Pronunciation Robustness benchmark. The conclusion is straightforward: Drama 3 ranks first in pronunciation robustness and is the top open-weight model in both speech arenas.
In the Provider Voice scenario, Drama 3 scores 1278 Elo with 1512 arena appearances, ranking 6th overall and 1st among open-weight models. It is 38 points above its own Drama 3 Preview at 1240 and 50 points above BreezeBlue's Breeze TTS 2 at 1228. In the Controlled Voice scenario, Drama 3 has 1085 Elo with 1958 appearances, ranking 12th overall and 1st among open-weight models, ahead of Ant Group's Ming-3.0-tts at 1068 and Preview at 1063.
On pronunciation robustness, Drama 3 scores 94.6% across 2,157 evaluated clips, ranking first among all models. Ming-3.0-tts and Drama 3 Preview both score 93.2%. This benchmark has human reviewers judge TTS performance on four categories of difficult text according to pre-agreed acceptable pronunciations.
Breaking out the subcategories, the conclusion is less even. Expanding Shorthand is 96.6%, ranking 1st. Contextually Appropriate is 96.5%, ranking 2nd, with Gemini 3.8 Flash TTS ahead at 97.9%. Preserving Exact Sequences is 87.1%, ranking 3rd, with HeyGen Voice Turbo leading at 91.8%. Standalone Terms is 91.8%, ranking 13th; its own Preview instead ranks 1st at 95.9%. An overall first place shows the model is more stable on difficult-to-pronounce text, but for standalone terms and exact sequences specifically, it still depends on the use case.
On speed, Drama 3 generates 66.6 characters per second. Preview is 60.7, and Murf's Speech Gen 2 is 63.0. ElevenLabs' Eleven v4 is 70.0, and v3 Conversational is 66.8, both a bit faster than Drama 3.
Price is another variable. Drama 3 costs $15 per million characters. Every model ranked above it on the Provider Voice leaderboard is more expensive: Gemini 3.8 Flash TTS at $16.5, Qwen-Audio-3.1-TTS-Plus at $19.3, Eleven v4 Turbo at $40, Sonic 3.6 at $49, and Eleven v4 at $80. One user converted by character volume and estimated that an hour of narration costs about $0.8, versus about $4.3 for Eleven v4. This number depends on text length and punctuation, but the direction of the price gap is clear.
Drama 3 is an open-weight model under the Fish Audio Community License. Its model card states 35B parameters with 3B activated per token, an MoE text backbone with hybrid attention, and an audio decoder using 10 codebooks. It supports multi-speaker output, voice cloning, and natural-language control; a single text can contain up to 16 speaker tags, and long text is automatically chunked with streaming output support. In the example, a voice is cloned from 3.9 seconds of reference audio, and inline tags such as [sighs] and [laughs] can place sighs and laughter at specific positions.
On deployment, Drama 3 could run on SGLang-Omni and vLLM-Omni on launch day. The Hugging Face model card provides streaming data on H200: single-stream first-audio-chunk p50 is 88.9 ms, with RTF 0.131. At concurrency 16, generation speed is still about 2.7x playback speed. At concurrency 64, average RTF is 0.801. Moving the vocoder to a second GPU increases throughput by about 72% at 256 concurrency and reduces median first-chunk latency by 37%. For people building real-time voice applications, these numbers are more direct than Elo.
On the Hugging Face page, Drama 3 already has 50 quantized versions, 7 fine-tunes, and 30 Spaces. The community is filling in the deployment and listening-demo layer.
One user said the 94.6% pronunciation score is what he cares about most; if the stress on one word is wrong, the whole video is ruined. Others asked about performance on Portuguese, Indian names, and Hinglish, and some wanted a language-by-language breakdown for Chinese polyphonic characters, numbers, and mixed Chinese-English reading. Artificial Analysis's overall score does not answer these questions. This is where the benefit of open weights comes in: running your own text through the model is more reliable than looking at the leaderboard.
Model and leaderboard links:
- Model: https://huggingface.co/fishaudio/drama-3/
- TTS leaderboard: https://artificialanalysis.ai/text-to-speech/leaderboard
- Controlled Voice leaderboard: https://artificialanalysis.ai/text-to-speech/leaderboard/controlled-voice
- Speech Arena: https://artificialanalysis.ai/text-to-speech/arena
Drama 3 pushes open-weight TTS one step further on both pronunciation and price. It does not win every subcategory, and deployment is not zero-cost. But for people building voice products, this model is already worth adding to the candidate list, especially in scenarios that are sensitive to both pronunciation errors and inference costs.
发布时间: 2026-10-30 06:35