Fish Audio S2.1 Pro: Expressive Real-Time Voice Model

Fish Audio S2.1 Pro is the model behind the player at the top of this page. It generates speech in real time with emotion control, reads scripts with multiple speakers, transcribes recordings, and powers voice cloning from a short sample. You can try every tab without paying; actual generation spends credits, and packs start at $9.

What is Fish Audio S2.1 Pro?

S2.1 Pro is the generation engine rather than a separate product. Everything you reach from the workspace — the text box, the clone workflow, the transcription tab, and the API — calls the same model family, which is why a take you hear in the browser is representative of what an integration produces.

That single-engine design has practical consequences:

• No re-learning between surfaces. Direction you write in the script behaves the same in the workspace and in an API call.
• Consistent voices. A voice chosen for one channel can be reused in another, including across languages.
• One credit system. Generation, transcription and cloning all spend the same credits, so budgeting does not require three spreadsheets.

How do you try S2.1 Pro without writing code?

The player has three tabs, and each one maps to a capability:

1. Text to Speech. Type a script, add emotion tags, choose a voice and generate. This is the fastest way to learn how the model handles your copy.
2. Voice Cloning. Record a short prompt and build a voice you can reuse — the consent and rights rules are covered in the cloning guide.
3. Speech to Text. Upload or record audio and read back a transcript with speaker separation and emotion tags; the speech to text guide goes further.

Nothing in the player is a mock-up. The scripts you type there can be moved to the workspace unchanged, and the same direction works from the API.

What does emotion control mean in practice?

Two layers do the work:

• Inline tags. Bracket markers such as [whispering], [excited], [pause] or [long pause] sit inside the sentence they affect. They are part of the text, so they survive a re-render.
• Described delivery. For longer passages, describe the read once — "warm, unhurried, as if explaining something to a colleague" — instead of tagging every line.

The difference matters most when a script has to do more than inform: a warning that should slow down, a punchline that needs a beat, a support answer that should sound reassuring rather than neutral. The text to speech guide lists the tags the engine reads.

How does it handle multiple speakers and languages?

Multispeaker dialogue is generated as one piece of audio rather than two files stitched together, which keeps interruptions and pauses in the right place. That is the difference between a conversation and a relay race, and it is what makes two-voice scripts usable in explainers, dramas and training material.

Language coverage is the other half. S2.1 Pro supports 30+ languages, and a voice is not locked to the language it was recorded in:

• The same narrator can carry an English series and its Japanese version.
• A brand voice can serve every locale instead of sounding like a different company in each market.
• A transcript in one language can be regenerated in another without re-casting.

What changes for long-form or production work?

Two upgrades matter once a project grows past a single take:

• Long-form workflow. Chapter-level control and paragraph-level regeneration keep fixes local; the Studio guide documents the pattern.
• API access. Included with the $99 Studio pack, together with the fastest queue and the highest concurrency for real-time streams. The API guide covers keys, credits and concurrency.

Real-time streaming is what makes the same model usable in a voice agent: audio starts playing before the paragraph finishes generating, which is what keeps a conversation feeling live.

What does using S2.1 Pro cost?

Every capability spends the same credits, at roughly one credit per 1,000 characters of speech:

Packs are one-time purchases, credits stay valid for 12 months, and there is no subscription to cancel. That structure suits evaluation: a single $9 pack is enough to decide whether the model fits a project before any long-term commitment.

Which workflows should you try first?

The model rewards starting with the workflow you already have, rather than the demo that looks most impressive.

• If you publish video or podcasts, start with text to speech on a real script and listen for pacing across a paragraph, not a sentence. The voice over guide covers the workflow.
• If you have recordings and no transcript, start with speech to text and check that speaker separation survives your audio quality.
• If a specific voice matters, start with cloning: record 15 seconds in the environment you will actually generate in — no room tone, no music, normal distance from the microphone.
• If you are building a product, start with the API and measure time to first audio before you optimise anything else; the voice agent guide sets out the latency budget.

In each case the decision is the same: does the output survive contact with the script you actually publish? That takes one paragraph and a few minutes, not a benchmark table.

Where should you start?

Start in the player on this page, with copy you actually intend to publish:

1. Paste a paragraph and generate it with two different voices.
2. Add one emotion tag to the line that carries the most meaning and regenerate.
3. Switch to the cloning tab if the project needs a specific voice.
4. Move to the workspace, the Studio workflow or the API once the direction is settled.

Ten minutes in the player answers more than a page of model specifications — and if you want the shortlist first, the voice library guide shows how to narrow two million voices down to three.

Frequently asked questions

What is Fish Audio S2.1 Pro?

It is the model behind the on-page player: real-time text to speech with emotion control, multispeaker dialogue, speech to text and voice cloning, all in the same engine you reach from the workspace.

Can I try S2.1 Pro before paying?

You can explore the player, the voices and the guides without paying, but generation itself spends credits. Audio generation starts with the $9 Starter pack — there is no free tier on this site.

What is emotion control?

Two layers: bracket tags such as [whispering] or [excited] inside the script, and natural-language descriptions of the delivery you want for a passage. Both travel with the script, so a re-render keeps the same direction.

Which languages does S2.1 Pro speak?

30+ languages including English, Japanese, Korean, Chinese, French, German, Arabic and Spanish. Voices are not tied to the language they were recorded in, so one voice can carry a series across markets.

Is S2.1 Pro available through the API?

Yes. API access is included in the $99 Studio pack, together with the highest concurrency for real-time streams.