| name | fish-audio-tts |
| description | Produce speech and clone voices with Fish Audio — its hosted TTS API (S2.1-Pro / S2-Pro / S1 model lineup, REST + WebSocket streaming, instant and persistent voice cloning) and its open-weight OpenAudio S1-mini / Fish-Speech models for self-hosting. Use this skill when an agent must generate narration or dialogue through Fish Audio, choose between Fish Audio's hosted models and open weights, clone a voice from reference audio, author emotion/tone/special markers for expressive delivery, estimate cost from UTF-8 bytes, wire real-time streaming for a voice agent, decide whether self-hosting beats the API, or review Fish Audio TTS output for production. Do not use it to pick a different provider — it covers Fish Audio specifically. |
Fish Audio text-to-speech and voice cloning
Fish Audio is a TTS and voice-cloning provider with two distinct product surfaces that share a lineage but differ in licensing and operation:
- Hosted API (
api.fish.audio) — the commercial service. Current lineup is branded S2.1-Pro / S2-Pro / S1. Paid per UTF-8 byte, streaming, cloning, and a free tier.
- Open weights — released under the OpenAudio brand (and earlier as Fish-Speech). Downloadable from Hugging Face and GitHub for self-hosting.
The single most important fact to get right: the hosted API and the open weights are not the same models, and their licenses differ. The full flagship weights are hosted-only; only smaller distilled weights are openly published, and those carry a non-commercial license. Never assume "Fish is open source, so I can self-host it commercially" — verify which artifact and which license apply. See Open weights and self-hosting below.
All model names, prices, endpoints, and limits below are volatile. Verification date for every dated claim in this document: 2026-07-10. Re-verify against docs.fish.audio before quoting these to a user as current.
Labels used throughout:
- [Doc] — stated in official Fish Audio / OpenAudio documentation, model cards, or their published technical reports.
- [Claim] — a first-party marketing or benchmark claim from Fish Audio/OpenAudio; treat as their assertion, not independently established.
- [Heuristic] — a production judgment from practice, not a documented guarantee.
When to use Fish Audio (and when not to)
Reach for Fish Audio when:
- You need low-cost, expressive multilingual TTS with strong voice cloning from short reference audio.
- You want fine-grained emotional/paralinguistic control through inline markers, not just SSML prosody.
- You need a real-time streaming voice for an agent (WebSocket, sub-second time-to-first-audio). [Claim]
- You want the option to self-host later, or to prototype locally on the open weights before committing to the API.
Prefer something else when:
- You need certified, contractual voice likeness rights for a specific celebrity/brand voice — Fish Audio does not pre-clear individual use cases and puts the consent burden on you. [Doc]
- Your deployment is commercial and you specifically want to run self-hosted open weights — the published weights are non-commercial-licensed, so commercial self-hosting needs the hosted API or a separate commercial license. [Doc]
- You need a provider with a formal, audited enterprise compliance posture (e.g., signed BAA/DPA guarantees on the free tier) — the free tier explicitly carries no such guarantees. [Doc]
- The languages you need fall outside the S1 supported set (see Languages) and you need documented, tested quality there.
This skill does not choose between providers. If the user is still deciding whether to use Fish Audio at all versus ElevenLabs, Cartesia, etc., surface the tradeoffs but recognize a provider-comparison skill governs that decision.
Hosted API
Endpoints [Doc, 2026-07-10]
- REST (batch/stream):
POST https://api.fish.audio/v1/tts
- WebSocket (bidirectional streaming):
wss://api.fish.audio/v1/tts/live
- Create voice model (cloning):
POST https://api.fish.audio/model
- ASR / transcription:
transcribe-1 model (separate endpoint; $0.36 per audio hour).
Required headers on TTS calls: Authorization: Bearer <FISH_API_KEY>, model: <model-name>, and Content-Type: application/json or application/msgpack. The model header — not a body field — selects the model. [Doc]
Model lineup [Doc, 2026-07-10]
| Model header value | Role | Marker syntax | Notes |
|---|
s2.1-pro | Recommended for production | square-bracket tags [...] | Improved quality, latency, throughput over S2-Pro. [Claim] |
s2.1-pro-free | Free tier, same weights as s2.1-pro | [...] | $0. No time-to-first-audio (TTFA) or DPA guarantees. For dev/prototyping/smaller use. [Doc] |
s2-pro | Previous-generation default | [...] | Multi-speaker, natural-language expression control. [Doc] |
s1 | Earlier generation | parenthesis tags (...) | Emotion markers via parentheses. [Doc] |
Critical marker-syntax difference [Heuristic, grounded in Doc]: S1 uses parentheses — (excited), (whispering). The S2 family uses square brackets — [excited], [whispers]. Markers written for the wrong model are often read aloud literally or ignored. Match marker syntax to the model header.
"OpenAudio S1" ≠ the hosted s1 model. OpenAudio S1 is the open-weight release brand (below). The hosted s1 is the corresponding hosted endpoint. Keep the two mentally separate when a user says "S1."
Core request parameters [Doc, 2026-07-10]
| Parameter | Type / range | Purpose |
|---|
text | string | Content to synthesize. In WebSocket mode, usually empty at start and streamed via text events. |
reference_id | string | ID of a saved voice (library or a cloned model). |
references | array | Instant cloning: raw reference audio + transcript inline, no persistent model. |
format | mp3 (default), wav, pcm, opus | Output container. |
mp3_bitrate | 64 / 128 / 192 | MP3 quality. |
sample_rate | Hz | For WAV output. |
speed | 0.5–2.0 | Speech rate. |
latency | normal (default) / balanced | balanced trades a little stability for lower latency. |
chunk_length | 100–300 (default 200) | Text batching size per synthesis chunk. |
normalize | boolean | Expands numbers/dates/abbreviations for natural reading. |
temperature | float | Lower = more deterministic/stable. |
top_p | float | Sampling diversity. |
repetition_penalty | >1.0 | Suppresses looping/repeated artifacts. |
max_new_tokens | int | Caps audio length per chunk. |
Format selection [Heuristic]: wav or pcm for downstream editing and telephony pipelines; opus for streaming to browsers/mobile; mp3 when you just need a portable file. pcm avoids container/decoder latency in real-time loops.
Streaming
Three generation modes in the Python SDK [Doc]:
tts.convert() — returns complete audio bytes (simplest; use for files/batch).tts.stream() — iterator of audio chunks (memory-efficient long-form).tts.stream_websocket() — real-time bidirectional streaming for agents.
WebSocket protocol [Doc, 2026-07-10] at /v1/tts/live, MessagePack-serialized:
- Client → server events:
start (config: format, chunk_length, reference_id, latency), text (append text as it's produced upstream, e.g. token-by-token from an LLM), flush (force immediate synthesis of buffered text — the key lever for low latency in interactive apps), stop (end of stream).
- Server → client events:
audio (a chunk of bytes in the requested format — concatenate all chunks for the full clip), finish (session end, with a reason).
Real-time heuristic [Heuristic]: stream the LLM's tokens into text events and issue flush at sentence or clause boundaries. Waiting for a whole paragraph before the first flush defeats the point of streaming. Fish's own material describes sub-500 ms end-to-end and ~450 ms perceived first-audio when streaming instead of waiting for the full file. [Claim]
Pricing and rate limits [Doc, 2026-07-10]
- All TTS models: $15.00 per 1,000,000 UTF-8 bytes.
s2.1-pro-free is $0.
- Reference scale: ~1M UTF-8 bytes ≈ 180,000 English words ≈ ~12 hours of speech. [Claim]
- Byte-based billing is the cost trap. Billing is by UTF-8 bytes, not characters. ASCII/Latin text is ~1 byte/char, but Chinese/Japanese/Korean/Arabic/Hindi characters are typically 3–4 bytes each. The same word count in Japanese costs several times more than in English. Estimate cost as
len(text.encode("utf-8")) / 1e6 * 15 USD, not from character count.
- ASR
transcribe-1: $0.36 per audio hour. Voice design voice-design-1: $0.01 per successful request.
- Concurrency tiers (by cumulative prepaid spend): < $100 → 5 concurrent; ≥ $100 → 15; ≥ $1,000 → 50; enterprise → custom.
- Free tier (
s2.1-pro-free) carries no TTFA or DPA guarantees. [Doc]
Voice cloning
Two paths [Doc, 2026-07-10]:
- Persistent model —
POST /model with type=tts, title, voices (one or more reference audio files), visibility (private/public), optional description. Returns a voice id; pass it as reference_id in TTS. Models move through created → trained; a "fast" train mode returns a usable voice immediately. enhance_audio_quality (default on) denoises and normalizes the reference.
- Instant cloning — skip model creation; send the reference audio + its transcript in the
references field of a TTS call. Best for one-off voices.
Reference-audio requirements [Doc + Heuristic]:
- Length: ≥ 10 seconds minimum; 30–60 seconds (up to 1–2 minutes) generally yields noticeably better similarity and stability. [Doc]
- Cleanliness: mono, single speaker, no background music, no reverb, no overlapping voices. Clean input dominates clone quality more than length. [Doc]
- Delivery: natural conversational speech, not exaggerated performance; include varied phonemes and intonation so the clone generalizes. [Doc]
- Formats:
.wav, .mp3, .m4a, .opus. [Doc]
- [Heuristic] If a clone sounds robotic or clips consonants, the reference — not the model — is usually the problem. Re-record a cleaner, longer sample before tuning generation parameters.
Emotion, tone, and paralinguistic markers
Fish Audio's differentiator is inline expressive control. Markers are written into the text itself. [Doc]
S1 (parentheses) [Doc, 2026-07-10]: 40+ markers across three groups.
- Emotion:
(angry), (sad), (excited), (surprised), (sarcastic), (joyful), (empathetic), and more.
- Tone:
(in a hurry tone), (shouting), (screaming), (whispering), (soft tone).
- Special/paralinguistic:
(laughing), (chuckling), (sobbing), (sighing), (panting), (crowd laughing).
S2 family (square brackets) [Doc, 2026-07-10]: a much larger open tag vocabulary (Fish cites 15,000+ tags) plus free-form natural-language descriptions. Documented examples include [pause], [short pause], [emphasis], [laughing], [chuckle], [sigh], [inhale], [exhale], [whisper], [shouting], [screaming], [loud], [low volume], [volume up], [volume down], [excited], [angry], [sad], [surprised], [singing], [echo], [with strong accent], [clearing throat], [audience laughter], and free-form forms like [whisper in a small voice]. [Doc/Claim — the 15,000 figure is a first-party claim.]
Marker heuristics [Heuristic]:
- One dominant emotion per sentence. Stacking
(angry)(sad)(excited) in one clause produces incoherent delivery.
- Place a marker immediately before the span it should color; it influences what follows, not the whole paragraph retroactively.
- Special markers like
(laughing) / [laughing] insert a non-verbal event — don't also spell out "haha" unless you want both.
- Free-form S2 tags are powerful but non-deterministic. If a specific effect must land every render, prefer a documented tag and lower
temperature.
- Always confirm the marker syntax matches the
model header (see the S1-vs-S2 warning above). This is the most common expressive-TTS failure with Fish.
Text preparation for natural delivery [Heuristic unless noted]
- Enable
normalize ([Doc]) so "$1,200", "Dr.", "3/4", and "2026" are expanded, or pre-expand them yourself for full control. Un-normalized numerics are a frequent source of wrong readings.
- Write punctuation for breath: commas and periods drive pacing more reliably than
[pause] for ordinary prose. Reserve pause markers for deliberate dramatic beats.
- Break very long inputs into sentence/paragraph units and let
chunk_length batch them; extremely long single chunks raise the odds of drift, repetition, or a runaway (laughing).
- Spell tricky proper nouns/acronyms phonetically if the model mispronounces them (e.g., "N. A. S. A." vs "NASA"), or add a light accent/emphasis tag.
- For dialogue with multiple characters, use separate cloned voices /
reference_ids per speaker rather than trying to shift one voice with markers alone.
Open weights and self-hosting
Fish Audio publishes open weights under the OpenAudio brand (successor branding to Fish-Speech). This is a separate track from the hosted API. [Doc, 2026-07-10]
What is actually open [Doc, 2026-07-10]
- OpenAudio S1-mini — 0.5B-parameter distilled model, publicly downloadable (Hugging Face
fishaudio/openaudio-s1-mini). Released alongside S1 on June 3, 2025.
- OpenAudio S1 (4B) — the full flagship is described as proprietary/hosted; the openly published artifact is the mini, not the full 4B S1.
- Fish-Speech 1.4 / 1.5 — earlier open releases (~0.5B class), still available.
- The GitHub project (
fishaudio/fish-speech) also describes a larger S2-Pro (4B) open weight in its README. [Doc] Confirm exactly which weight file and license you are downloading before relying on it.
Licensing — read carefully [Doc, 2026-07-10]
- OpenAudio S1-mini weights: CC-BY-NC-SA-4.0 (per the Hugging Face model card). CC-BY-NC-SA means non-commercial, attribution, share-alike. You cannot use these weights in a commercial product without a separate commercial license.
- The OpenAudio GitHub code is Apache-2.0; its weights are CC-BY-NC-SA-4.0. [Doc]
- The
fishaudio/fish-speech repository's README states code and weights under the Fish Audio Research License (with an explicit "we will act against violations" notice). [Doc]
- [Heuristic / caution] Third-party review sites sometimes claim these weights are "MIT-licensed." That contradicts the primary model card and repo license and is unverified — do not rely on it. When commercial use is on the table, treat the open weights as non-commercial and route commercial workloads through the paid hosted API (or negotiate a commercial license).
Hardware [Doc + practitioner reports, 2026-07-10]
- Inference minimum: ~12 GB VRAM (documented self-hosting minimum). [Doc]
- S1-mini specifically: ~3.5 GB download, ~5 GB VRAM to run — fits comfortably on 8–12 GB consumer GPUs. [Practitioner reports — labeled secondary]
- S2-Pro (4B): substantially heavier; ~24 GB VRAM recommended for comfortable inference. [Practitioner reports — labeled secondary]
- Real-time factor scales with the GPU: e.g., ~100 ms TTFA and RTF ~0.195 on a single NVIDIA H200 in Fish's own S2 report [Claim]; consumer cards (RTX 3060/4090) are usable but slower. [Secondary]
- Stack: Linux or WSL; audio deps (
ffmpeg, libsox, portaudio); install via conda or uv; vLLM / SGLang backends for throughput. [Doc]
When self-hosting beats the API [Heuristic]
- Self-host when: you need data to stay on-prem; you run very high, steady volume where $15/M bytes exceeds amortized GPU cost; you need offline/air-gapped operation; or you are doing research/personal/non-commercial work where the CC-BY-NC-SA license is fine.
- Use the hosted API when: you need the full-quality flagship weights (not published openly); you need commercial rights without a bespoke license; you want managed scaling, streaming SLAs, and the newest S2.1 quality; or your volume is low/bursty enough that GPU idle time makes self-hosting more expensive than per-byte billing.
Benchmark standing (disclosed methods, labeled)
Treat these as vendor and third-party claims with stated methods, not as neutral ground truth. All dated 2026-07-10 as read.
- OpenAudio S1 — TTS-Arena-V2 (Hugging Face), human subjective ELO: ranked #1 as of June 3, 2025. [Claim, first-party announcement; method = crowd human preference.]
- OpenAudio S1 — Seed-TTS-Eval (transcription-based WER/CER): S1 WER 0.008 / CER 0.004; S1-mini WER 0.011 / CER 0.005 (English). [Claim, first-party; method = ASR-scored intelligibility.]
- Artificial Analysis speech arena: OpenAudio S1 ELO ≈ 1,074. [Third-party, disclosed-arena method.]
- Fish Audio S2 technical report: Seed-TTS-Eval WER 0.54% (Chinese) / 0.99% (English); EmergentTTS-Eval 81.88% win rate; Audio Turing Test posterior mean 0.515; RTF 0.195 and TTFA ~100 ms on H200. [Claim, first-party technical report; EmergentTTS-Eval is a NeurIPS'25 benchmark using an audio-LLM judge across six prosodic/expressive scenarios.]
How to use benchmarks in production advice [Heuristic]: low WER/CER means intelligible, not expressive or preferred. Arena ELO reflects listener preference on the arena's prompt mix, which may not match your domain (long-form audiobook vs. short agent turns). Always validate on your own scripts, voices, and languages before committing.
Voice-clone consent and misuse safeguards [Doc, 2026-07-10]
Non-negotiable, and the agent should enforce these when a user asks to clone a voice:
- Clone only voices you have the right to use: your own voice, a voice whose owner has given explicit consent, or a properly licensed voice. Fish Audio's terms put this obligation on the user. [Doc]
- The user is responsible for having the rights, consents, and disclosures required, and for complying with local law on name, likeness, and AI-generated media. [Doc]
- Fish Audio does not pre-clear individual use cases and may remove content or accounts that violate its terms or applicable law, at its discretion and without notice. [Doc]
- Do not produce deceptive impersonation, non-consensual clones of real people (public figures included), fraudulent/scam audio, or content that hides its synthetic nature where disclosure is required.
- [Heuristic] If a user asks to clone a named celebrity or a private individual without evidence of consent/rights, decline the clone and offer a licensed stock voice or a synthetic non-matching voice instead. Recommend disclosing AI generation to end listeners.
- Cloned voice models can be deleted from the account; treat reference audio and clones as sensitive data (see privacy note).
Failure modes and repair [Heuristic unless noted]
| Symptom | Likely cause | Fix |
|---|
| Markers read aloud literally ("open paren excited close paren") | Wrong marker syntax for the model | Match (...) to s1, [...] to S2 family |
| Robotic / low-similarity clone | Noisy, short, or multi-speaker reference | Re-record clean mono ≥30 s; keep enhance_audio_quality on |
| Repetition / looping / runaway laughter | Long chunk, high temperature, or stacked markers | Shorten chunks, raise repetition_penalty, lower temperature, one marker per span |
| Numbers/dates/acronyms misread | Normalization off | Set normalize=true or pre-expand text |
| Bill far higher than expected | Non-Latin script billed at 3–4 bytes/char | Estimate from utf-8 byte length, not char count |
| High first-audio latency in an agent | Waiting for full text before synthesis | Use WebSocket, stream text, flush at clause boundaries, format=pcm, latency=balanced |
| Commercial/legal exposure from self-host | Used CC-BY-NC-SA weights commercially | Move to hosted API or obtain a commercial license |
Production review checklist [Heuristic]: intelligibility (no dropped/garbled words), correct pronunciation of names/numbers, appropriate and non-literal marker rendering, consistent voice identity across chunks, natural pacing/breaths, no repetition or clipping at chunk seams, target loudness, and — for clones — documented consent on file.
Complete examples (illustrative, not mandatory formulas)
Example A — Expressive narration with S2.1-Pro (REST, cost-aware)
Intent: a 2-sentence dramatic voiceover with a whispered aside, delivered as a WAV for editing; estimate cost first.
Model: s2.1-pro (square-bracket markers).
TEXT='The vault was empty. [whisper] Someone had been here first.'
python -c "print(len(open('/dev/stdin').read().encode())/1e6*15)" <<< "$TEXT"
curl -s https://api.fish.audio/v1/tts \
-H "Authorization: Bearer $FISH_API_KEY" \
-H "model: s2.1-pro" \
-H "Content-Type: application/json" \
-d '{
"text": "The vault was empty. [whisper] Someone had been here first.",
"reference_id": "<your-voice-id>",
"format": "wav",
"sample_rate": 44100,
"normalize": true,
"temperature": 0.7,
"repetition_penalty": 1.2
}' --output vault.wav
Why structured this way: brackets match the S2 model; the whisper marker precedes only the aside; normalize guards stray numerics; modest temperature with a repetition penalty keeps the dramatic line stable. Likely failure: if the voice were an s1 clone you'd need (whisper) instead. Variation: switch to s2.1-pro-free for a zero-cost test render before spending on the paid model.
Example B — Persistent voice clone, then reuse (Python SDK)
Intent: clone a consented narrator once, reuse across many lines.
from fishaudio import FishAudio
client = FishAudio()
voice = client.voices.create(
title="Narrator - Priya (consented 2026-07-10)",
voices=[open("priya_45s_clean.wav", "rb").read()],
description="Studio sample, single speaker, no music",
visibility="private",
)
audio = client.tts.convert(
text="Chapter one. (soft tone) It began on a grey morning.",
reference_id=voice.id,
)
open("ch1.mp3", "wb").write(audio)
Why: the reference is long, clean, mono, and consented; the clone is private; markers match the model family. Failure: a public reference sample with background music degrades every downstream render. Variation: for a single throwaway line, use references=[...] instant cloning instead of creating a persistent model.
Example C — Real-time voice agent (WebSocket)
Intent: speak an LLM's tokens as they arrive with minimal first-audio latency.
Pattern (pseudocode over wss://api.fish.audio/v1/tts/live, MessagePack):
send start { format:"pcm", latency:"balanced", chunk_length:200, reference_id:"..." }
for token in llm_stream:
send text { text: token }
if token ends a clause: send flush {}
send stop {}
# concatenate every server `audio` event's bytes until `finish`
Why: pcm avoids decode latency; flush at clause boundaries emits audio without waiting for the full reply; balanced shaves latency. Failure: flushing only at the end reintroduces the ~1 s+ full-file wait. Variation: raise chunk_length for smoother prosody on long, non-interactive narration where latency doesn't matter.
Example D — Self-host decision
Intent: user has 40 hours/month of internal, non-commercial research narration and one RTX 4090.
Decision: OpenAudio S1-mini self-hosted is reasonable — fits the 4090, non-commercial use satisfies CC-BY-NC-SA, and steady volume amortizes the GPU. [Heuristic] But if that narration ships in a paid product, self-hosting the NC weights is a license violation → use the hosted API instead. Confirm the actual downloaded weight's license before deploying.
Privacy note [Heuristic]
Reference audio for cloning is biometric-adjacent personal data. Keep clone models private, store consent records with the voice, avoid uploading third-party voices you cannot substantiate rights for, and delete clones and source audio when the project ends. Do not send user PII inside TTS text to a third-party API without the user's awareness.
Primary sources (verified 2026-07-10)
