Fish Audio S1
Audio Deprecated Full comparison ↗ fish-audio/s1 · by Fish Audio
· neural-tts
Pricing — 1 offering(s)
About this offering: Fish Audio S1: a 4B-parameter TTS model with 13 languages and 64+ emotion and tone markers. Fish Audio has deprecated it and will retire it on 2026-12-31; it recommends S2.1 Pro for new projects (Fish Audio's models overview, checked 2026-10-10). After that, requests for s1 are routed to S2.1 Pro and billed as S2.1 Pro, at the same price, so plan new work on S2.1 Pro.
Text characters
- $15.00 / 1M characters
Current
2026-07-30 → 2026-12-30
Fish Audio's native billing unit is UTF-8 BYTES, not characters — the pricing page states "$15.00 / M UTF-8…
Fish Audio's native billing unit is UTF-8 BYTES, not characters — the pricing page states "$15.00 / M UTF-8 bytes". For English and other Latin-script text 1 character = 1 byte, so the rate is equivalent; for CJK and other multi-byte scripts, actual cost per character can run up to 3-4x this rate. Recorded here as per_1m_characters (closest schema unit) since the registry has no byte-based unit; The price's start date is the date Modelglass first recorded it, not a confirmed price-change date.
Showing the active price and any recorded history. Full pricing history is available via the paid API — see API docs.
See how Fish Audio S1 fits into a cost-aware routing setup
See how →Capability profile
How Fish Audio S1 rates across core capability dimensions, with the task-level evidence behind each rating.
Fish Audio reports a 0.8% word error rate and 0.4% character error rate. In the Artificial Analysis Text to Speech Arena (checked 2026-08-19), it's listed as "OpenAudio S1" (Fish Audio's research brand) at #42 of 98 voices (Elo 1079).
13 languages — narrower than S2 Pro/S2.1 Pro (80+/83) and below ElevenLabs Multilingual v2 (29) or Eleven v3 (70+).
Voice cloning from as little as 10-15 seconds of reference audio, plus access to Fish Audio's community voice library (2M+ voices claimed platform-wide, not S1-specific).
No S1-specific latency figures published. S2 Pro/S2.1 Pro carry an explicit 100ms time-to-first-audio guarantee that S1 does not — Fish Audio's docs don't state whether S1 shares comparable latency.
Zero-shot voice cloning from ~10-15 seconds of reference audio, per Fish Audio's platform-wide cloning claim.
Benchmarks
| Benchmark | Score | Config | Source |
|---|---|---|---|
| Word error rate (vendor-reported) | 0.8 % | Fish Audio's own reported WER. Not independently re-measured by Modelglass. | source ↗ |
| TTS-Arena2 ranking (independent, community leaderboard) | — | — | — |
Operator guidance
Deprecated: Fish Audio retires S1 on December 31, 2026, after which s1 requests are served and billed as S2.1 Pro at the same price, so don't build new long-term work on it. Fish Audio's own docs label S1 "Previous Model" and recommend s2.1-pro for all new projects — S1 is kept available only "for existing integrations," priced identically to S2 Pro/S2.1 Pro. Use S2.1 Pro instead for anything new; S1 is documented here for operators already integrated against it.
Use cases
- Existing integrations already built on S1 (Fish Audio's own guidance — new projects should use S2.1 Pro instead)
- Cost-sensitive TTS with strong naturalness at S2/S2.1 Pro's same price point, until S1's 2026-12-31 retirement
- Expressive narration via 64+ emotion/tone markers using (parenthesis) syntax
Limitations
- Fish Audio itself recommends against using S1 for new projects — kept only for existing integrations
- Narrower language support (13) than S2 Pro/S2.1 Pro (80+/83) or ElevenLabs' models
- No published time-to-first-audio guarantee, unlike S2 Pro/S2.1 Pro's 100ms SLA
- Fish Audio's native billing unit is UTF-8 bytes, not characters — see registry pricing notes