Fish Audio Drama 3 Release: 35B MoE TTS Deep Dive

Fish Audio Drama 3 Release: Deep Dive Into 35B-A3B Open Weights and Word-Level Audio Inpainting

Fish Audio Drama 3 shipped on October 13, 2026, about a day before this piece was written. It did not arrive as a closed API, a waitlist, or a demo reel. It came out as a downloadable checkpoint, a ComfyUI node, and a day-zero serving recipe, all within the same afternoon. Here's what's actually in it, and what still has to be checked.

TLDR Fish Audio Drama 3 is a 35B-parameter mixture-of-experts text-to-speech model with 3B parameters active per token. It ships open weights under the Fish Audio Community License. Its main features are plain-language performance direction, mid-sentence emotion shifts, multi-speaker scenes in one pass, and Audio Inpainting for single-word fixes. The hosted API ID is reportedly drama-3-preview, with one outlet citing $15 per 1M UTF-8 bytes. No benchmark figures have been verified yet.

Key Takeaways

• Architecture: 35B total parameters, 3B active per token, hybrid full and linear attention, and a 10-codebook audio decoder.
• Open weights: The checkpoint is downloadable under a custom community license, not Apache or MIT.
• Control surface: Inline tags such as [sigh] and <whisper>…</whisper>, plus free-text direction like "gentler, and a bit sadder."
• Scene scale: Up to 16 speaker tags in a single text, rendered in one generation.
• Open questions: No verified benchmarks, no latency numbers, conflicting language counts, and unconfirmed pricing.

What Actually Shipped

The launch had several parts on the same day. Fish Audio's press release went out at 9:15 PM GMT+8 on October 13. It describes a model that lets users "direct a voice performance instead of simply generating one." Community posts followed within hours.

Alvaro Cintas summarized the pitch: write "[crowd laughing]" and the model performs it, shift emotion mid-sentence with no retake, and fix a single word without regenerating the whole take. He noted the model can be tried on the web or through the API from launch day.

ComfyUI announced Drama 3 as a Partner Node. Per that post, it generates "the whole multi-character scene at once" and pairs with Fish Audio TTS, voice cloning, and transcription. It was the most-seen post in this cycle at 5,044 views and 83 likes.

The weights are the bigger news for engineers. Launch coverage reports a Hugging Face repository under fishaudio/drama-3, with a README, serving recipes, and a linked technical report. Charly Wargnier flagged it plainly: "Open-weight, too." @kimmonismus added that the weights are available under Fish Audio's Community License.

The SGLang-Omni cookbook page for Drama 3 is the most technical public document. It shows a two-line setup: hf download fishaudio/drama-3, then sgl-omni serve --model-path fishaudio/drama-3 --port 8000. It exposes an OpenAI-style /v1/audio/speech endpoint for plain TTS, voice cloning from a reference clip, and PCM streaming.

The model also has a history. Launch coverage says Drama 3 was first teased in September behind a "comment DRAMA for a key" gate. October 13 marks its exit from that gated preview.

The Architecture: 35B-A3B and the Codebook-MTP Decoder

The SGLang-Omni cookbook lists five concrete specs:

Component	Detail
Total parameters	35B
Active per token	3B (MoE text backbone)
Attention	Hybrid full + linear
Audio decoder	Codebook-MTP, 10 codebooks decoded causally
Precision	FP8 MoE experts, remaining weights in bf16
Speaker tags	<|speaker:0|> to <|speaker:15|> (16 characters)
Long text	Chunked automatically, streamable per chunk

The comparison with Fish Audio's previous flagship is useful. Launch coverage lists S2-Pro at 4B parameters, and it remains available. At 35B total parameters, Drama 3 is roughly 9× the size of the 4B S2-Pro, yet it activates fewer parameters per token (3B) than S2-Pro's full 4B. That trade is a common one. Sparse experts add capacity, while per-token compute stays near the older model's.

The Codebook-MTP Audio Decoder is Drama 3's audio output stage, which decodes 10 codebooks causally across the whole sequence rather than one token stream at a time. "MTP" suggests multi-token prediction, but the cookbook does not spell out the acronym. Treat that reading as plausible, not confirmed.

FP8 experts matter for deployment. They should shrink the memory footprint of the largest part of the model. The cookbook's default command puts preprocessing, the engine, and the vocoder on GPU 0. That implies single-GPU serving is the intended baseline. No VRAM figure appeared in the public sources reviewed, though.

Per launch coverage, the fish-speech repository uses one interface for Drama 3 and S2-Pro and detects the model type from the checkpoint config. Teams already running S2-Pro may find migration fairly cheap. That is an inference from the docs, not a tested result.

The Control Surface: Directing Instead of Generating

Fish Audio's framing is that voice AI should move from generation to direction. The press release lists four capabilities. The README reportedly lists seven. These are the named features that matter most:

Natural-Language Direction is the ability to describe a delivery in plain text, for example "[warm and relaxed, like talking to an old friend]", and to change it mid-line. A sentence can start calm and turn panicked without a new take.

Point Tags and Paired Tags are the two inline syntaxes. Square-bracket tags such as [sigh] or [laughs] fire at their position in the script. Paired tags such as <whisper>Go on without me.</whisper> limit an effect to the enclosed words. That is how Cintas's "whisper or slow down only the words you select" works in practice.

Audio Inpainting is Drama 3's ability to regenerate a selected word or segment while conditioning on the audio on both sides, so voice, mood, and rhythm stay intact. Shijia Liao, Fish Audio co-founder and chief scientist, called it "one of the hardest parts of the model to get right."

Multi-Turn Director Revision lets a user guide the previous take with a follow-up note, such as "gentler, and a bit sadder," instead of rewriting the prompt from scratch.

Multi-Speaker Scenes render several characters in one generation. According to the press release, each character keeps its own voice and responds to what the other just said.

Auto Tag proposes an initial set of performance directions for unannotated dialogue. It could matter most for teams converting existing scripts in bulk.

The Fish Audio homepage shows the tag vocabulary in its demo widget. It has 8 emotion tags ([angry], [sad], [embarrassed], [emphasis], [whispering], [soft], [breathy], [excited]) and 14 special tags, including [crowd laughing], [clear throat], and [long pause]. That is 22 preset tags before any free-text direction.

Why This Matters

The first reason is licensing. Fish Audio's previous premium tier, S2.1 Pro, reportedly stayed behind the paid API. Drama 3 reverses that for the flagship. Aditi pointed to that lineage: "Fish Speech was out there for everyone." Open checkpoints for expressive, directable TTS are still rare. A flagship-tier release changes what self-hosting teams can evaluate.

The second reason is workflow economics. Audiobook and dubbing pipelines lose time when one wrong word forces a full regeneration. CEO Rissa Cao made this argument directly in the launch release: "one wrong word shouldn't mean regenerating a whole chapter." If Audio Inpainting works as described, it removes a manual splicing step. That benefit could compound over hours of narration.

The third reason is attention. @kimmonismus observed that "text-to-speech models receive little attention, yet they are currently experiencing similar breakthroughs." His example is the line "I didn't say that," which changes meaning depending on which word is stressed. Per-word emphasis control targets exactly that kind of ambiguity.

The business context is reported, not audited. One newsletter cites a $52M seed round and more than 15M company-reported users. Fish Audio's own homepage says "Trusted by 8M+ users." The two figures may measure different things, but they do not match.

Fish Audio Drama 3 vs ElevenLabs: What the Reports Show

ElevenLabs is the competitor named most often in the launch discussion. A widely shared Facebook video opened with "Eleven Labs must be crying in their soup today." A preview-era YouTube review framed Drama 3 and S2.1 Pro as a possible "best ElevenLabs alternative." The comparison mostly comes from impressions, and that should be said plainly.

• Weights and license: Drama 3 ships open weights under the Fish Audio Community License. The available sources say nothing about ElevenLabs' weights or licensing, so this side is unverified. No public number from either lab yet.
• Hosted API pricing: One outlet reports $15 per 1M UTF-8 bytes for drama-3-preview. A second outlet explicitly says it could not confirm that figure. ElevenLabs pricing is unverified in these sources. No public number from either lab yet.
• Published benchmarks: Unverified. No public number from either lab yet. A Drama 3 technical report reportedly exists, but none of its figures appeared in the sources reviewed.
• Voice quality and emotional nuance: The only head-to-head claim is a testimonial on Fish Audio's own homepage. It says Fish Audio "clearly outperformed in voice authenticity and emotional nuance." A vendor-hosted quote is not a measured test.
• Multi-speaker dialogue: Drama 3 documents up to 16 speakers per text. The sources reviewed have no matching ElevenLabs spec.

A short YouTube review from the Aira Tech channel (199 views at the time of capture) tests emotion control and voice cloning. It was posted about 6 days before the October 13 launch, so it likely covers the preview build rather than the open-weight release.

Teams that want a fair comparison should run the same script through both systems. ElevenLabs V3 is a natural baseline for multi-speaker dialogue tests.

On the limited evidence so far, Drama 3 appears to lead on openness and editability. Whether it leads on raw voice quality has not been measured publicly.

What We Know vs. What We Don't

What we know

• Drama 3 is a 35B-parameter MoE model with 3B active per token, hybrid full and linear attention, and a 10-codebook Codebook-MTP decoder, per the SGLang-Omni cookbook.
• Weights are downloadable under the Fish Audio Community License, which is not Apache or MIT, per launch coverage and @kimmonismus.
• One text can carry up to 16 speaker tags (speaker:0 to speaker:15).
• Audio Inpainting regenerates a selected segment while conditioning on the surrounding audio, per the press release.
• The hosted API identifier is reportedly drama-3-preview.
• Drama 3 runs as a ComfyUI Partner Node and has day-zero SGLang-Omni support.

What we don't know

• Benchmarks: No verified figures. The technical report's numbers have not been independently checked.
• Pricing: $15 per 1M UTF-8 bytes is reported by one outlet and explicitly unconfirmed by another.
• Language count: @kimmonismus cites 83 languages. The press release names English, Chinese, Japanese, and Korean for phoneme control. Fish Audio's homepage says "30+ languages," possibly platform-wide.
• Latency: The homepage calls Drama 3 a "real-time voice model," but no latency or throughput figures are public.
• Long-form coherence: The press release says Drama 3 "holds one performance across the full script." The cookbook says long text "is chunked automatically." How context carries across chunks is unclear.
• License terms: The commercial limits of the Community License were not detailed in the sources reviewed.

What this tells us: the architecture, the control syntax, and the distribution channels are documented. What it doesn't tell us: how good the model is against a fixed test set, what it costs at scale, and how it behaves on chapter-length input.

How to Evaluate It Yourself

The fastest route is the cookbook's local server. Its voice-cloning example uses a 3.9-second reference clip with a known transcript. Fish Audio's homepage separately claims cloning "with perfect fidelity in 15 seconds." Test both lengths. A gap between 3.9 s and 15 s references would tell you how much reference audio your product needs.

Build a four-part test script:

1. Inpainting seam test. Generate a 60-second paragraph, swap one mid-sentence word, then listen at the boundaries for pitch or room-tone jumps.
2. Emphasis test. Render "I didn't say that" with stress on each of its four words. Check whether the meaning actually changes.
3. Scene test. Write a scene with at least 4 speakers and long gaps between each character's lines. Check whether voices drift by the end.
4. Chunk-boundary test. Feed a script long enough to trigger automatic chunking. Listen for resets in energy or pacing where chunks meet.

Record wall-clock time per second of audio on your own GPU. Until Fish Audio or a third party publishes latency figures, that number is the only one you can trust.

What Builders Should Do Today

Read the Community License first. It is a custom license. Its terms for commercial use, redistribution, and fine-tuning decide whether self-hosting is viable for your product.

Pin the version. The hosted ID still carries a -preview suffix. Behavior and pricing may change. Log the exact checkpoint revision you test against.

Keep S2-Pro in the comparison. At 4B parameters it is still available, and it shares one interface with Drama 3 in the fish-speech repository. For simple narration with no direction, the older model may be enough.

The Week Ahead

Watch for independent numbers from the technical report: word error rate, speaker similarity, and any preference studies. Run your own inpainting seam test before relying on the single-word-edit claim in production. Check whether Fish Audio publishes a pricing page that confirms or replaces the reported $15 per 1M UTF-8 bytes, and whether the 83-language figure appears in official docs.

Try related models on Kie.ai: Gemini 3.8 Flash TTS, Gemini 3.8 Flash-Lite TTS and Gemini 3.1 Flash TTS.