Fish Audio 發佈 Drama 3，支援 AI 語音指導及單字級音訊修補

A line of dialogue can contain all the right words and still deliver the wrong performance. The pause arrives too early. A character sounds confident when the scene calls for hesitation. One mispronounced word breaks an otherwise convincing narration.

Fish Audio is targeting those production problems with Drama 3, its new text-to-speech model announced on October 13. The company says creators can direct delivery in ordinary language, generate scenes with multiple speakers, and repair a selected segment of synthesized audio without regenerating the surrounding performance.

The most consequential feature may be the smallest edit: changing a single word while preserving the voice, mood, and rhythm on either side. For teams building audiobooks, dubbed content, games, or audio dramas, the value of a synthetic voice increasingly depends on how easily it can be revised.

Giving a Voice Performance Direction

Drama 3 accepts descriptions of tone, pacing, and character, including instructions that change delivery within a sentence. A passage can begin calmly and become panicked, according to the announcement, rather than requiring a new generation for each emotional section.

“Voice AI has gotten very good at generating speech. The next step is directing a performance,” said Rissa Cao, Fish Audio’s co-founder and CEO, in the launch announcement.

The company’s public Drama 3 preview page provides a more concrete view of the controls. Users can place directions, sounds, and pauses in square brackets, or apply paired tags to selected words. That lets a writer scope an effect such as a whisper, emphasis, or slower delivery to a particular phrase. An Auto Tag feature can also suggest initial directions for a script, which creators can then revise.

The creative implication is straightforward: a script becomes both dialogue and a set of performance instructions. Writers can work on what a character means, where a breath belongs, and which phrase deserves emphasis in the same editable text.

Audio Inpainting: Fixing the Moment That Fails

Fish Audio describes Drama 3’s audio inpainting as a way to select a portion of an existing performance, update the instructions, and regenerate that segment while retaining the surrounding audio. Co-founder and chief scientist Shijia Liao says the model uses the sound on both sides of an edited word to make the replacement fit.

That is a different task from generating an isolated replacement line. The new audio has to connect to an established performance: its timing, intonation, and vocal identity all need to remain believable at the boundaries of the edit.

Consider a fictional audiobook production in which a narrator delivers a paragraph correctly but stresses one name awkwardly. A targeted repair could preserve the successful delivery of the rest of the paragraph. The same approach could help a game team revise a short piece of dialogue after a script change.

These are potential workflow benefits, rather than measured productivity gains. The release does not provide an independent evaluation of how reliably single-word repairs preserve surrounding audio across voices, accents, or difficult transitions.

Scenes, Characters, and Multilingual Consistency

Drama 3 also supports generating a complete multi-speaker scene in one pass. Fish Audio says each character retains a distinct voice and responds in the context of the preceding dialogue. The launch emphasizes consistent characters across long scripts, language changes within a scene, and phoneme-level pronunciation control in languages including English, Chinese, Japanese, and Korean.

For dialogue, continuity involves more than recognizing which person is speaking. A response needs to sound like it follows the previous line, rather than like an unrelated recording pasted into a sequence. For localization, pronunciation and characterization must also survive the move between languages.

The company’s claims address both problems, but the practical test will be the finished scene: listeners should not hear abrupt changes in identity or delivery when speakers switch, a passage is edited, or the script changes language.

Where Drama 3 Fits in Fish Audio’s Technology

The direction-first approach builds on a research path Fish Audio has already described. Its earlier S2 research release supported natural-language instructions embedded at specific word or phrase positions. The company described a dual-autoregressive architecture that separates prediction along the audio timeline from generation of finer acoustic detail, alongside reinforcement-learning alignment for intelligibility, instruction adherence, and voice similarity.

That earlier work supplies useful context for Fish Audio’s focus on controllable speech. It does not establish Drama 3’s exact architecture: the new announcement refers to an architectural change but does not publish enough detail to equate the two systems or transfer S2’s benchmark results to Drama 3.

For developers, the text-to-speech API documentation lists Drama 3 under the model identifier drama-3-preview. Multi-speaker requests use a list of voice identifiers and speaker markers within the text. The endpoint also exposes controls for speed, volume, sampling, and output formats including WAV, PCM, MP3, and Opus.

The public product page and API documentation still describe Drama 3 as a preview. Fish Audio advises that results vary between takes and that the model will continue to change. Production teams should therefore distinguish the launch’s feature claims from the consistency they observe in their own material.

A Production Tool Is Judged by Its Revisions

Fish Audio says it now serves more than 15 million users and has reached $21 million in annual recurring revenue. Those company-reported figures place Drama 3 within a broader platform spanning speech generation, transcription, voice cloning, and creative tools.

For the launch, the company says API usage will be discounted by 80% for the first 30 days, with three free months for qualifying enterprises switching voice providers. Readers can check Fish Audio’s pricing page for current plans and offer details.

Drama 3’s proposition is strongest at the point where generation meets revision. A compelling first take is useful; keeping its successful moments while fixing the weak ones is what makes a voice model easier to fit into a production workflow. The launch puts that capability at the center of Fish Audio’s next chapter.