Fish Audio
Clone a voice from a 10-second clip and generate lifelike speech via API or playground, now on the S2.1 Pro model with emotion tags, fresh off a $52M seed round.
Alternatives
Fit Score: use-case reference
How this is scoredFish Audio scored 1 to 5 on five dimensions for the named use cases below. Part of our catalog-wide use-case reference, not a product-quality review rating.
A developer who wants production-grade voice cloning or speech synthesis via API without building on a consumer-facing abstraction
4.2/5- Capability
- 5/5
- Reliability
- 4/5
- Value
- 4/5
- Ease
- 4/5
- Compatibility
- 4/5
Fish Audio is the same TTS infrastructure already running in production inside HeyGen and Retell, now available directly as a standalone API and playground, with voice cloning that works from a 10-second clip and no minimum API commitment.
A team wanting to self-host the open-weight models without an enterprise-level engagement
2.0/5- Capability
- 2/5
- Reliability
- 3/5
- Value
- 2/5
- Ease
- 1/5
- Compatibility
- 2/5
Cons are direct that self-hosting the open-weight models is an enterprise-level engagement, not a quick setup, and commercial voice cloning rights require a paid plan rather than the free tier.
Editorial judgment grounded in this tool's own published pricing, features, and our existing pros and cons, not a certification or a measured benchmark. Product quality is assessed separately under our review methodology.
Overview
Not yet reviewedProduct quality is assessed from original evidence. How reviews work.
Fish Audio builds the text-to-speech models that quietly power voice features inside HeyGen, Retell, Sanas, and other AI products, and it also ships that same technology as a standalone playground and API. Its newer flagship model, S2.1 Pro, adds in-text emotion tags, marking a line as angry, whispering, or laughing, ultra-low latency, and support for 30-plus languages, built on top of the S1 lineage that was trained on more than 2 million hours of audio and hit a 0.8% word error rate on the Seed TTS benchmark. Voice cloning needs as little as 10 to 30 seconds of reference audio with no separate training step.
A $52M seed round in late July 2026 confirmed real traction: 8M-plus users and roughly $21M in annualized revenue at seed stage, unusually strong numbers for a company still raising its first institutional round. The API is pay-as-you-go from the first call, priced at $15 per million UTF-8 bytes of input text with no seat fees or minimum commitment, and a free-tier model variant exists for testing and prototyping without production guarantees. For teams that need to self-host, Fish Audio's core models ship as open weights under a paid commercial license (three of its five released models are fully open-source), deployable in a private VPC, on-prem, or air-gapped environment.
The consumer-facing playground adds credit-based subscription plans with unlimited public voice slots on every tier, making it accessible for solo creators as well as engineering teams shipping voice into their own product.
The source of the legacy catalog rating and review count has not been substantiated. Neither is presented here as a verified review rating or included in review structured data.
Workflow Fit Scores are separate, use-case-specific editorial judgments, not user-review averages. Read the Fit Score methodology.
- • Same model quality already trusted in production by HeyGen and Retell
- • Genuinely usable free tier plus API pricing with no minimum commitment
- • Voice cloning quality holds up well even from short reference samples
- • Credit-to-minute math on the consumer plans takes a minute to understand
- • Best commercial voice cloning rights require a paid plan, not the free tier
- • Self-hosting the open-weight models is an enterprise-level engagement, not a quick setup
Key Features
- Zero-shot voice cloning from 10-30 seconds of reference audio
- S1 flagship model with open-domain emotion and tone markers
- Pay-as-you-go API with WebSocket streaming and Python/TypeScript SDKs
- Open-weight self-hosting option under a commercial license
- Multilingual, cross-lingual speech with no phoneme dependency
Trust & Data
Verified 2026-09
- Training on your data
- Not publishedFish Audio's own privacy policy only vaguely lists research and internal analytics as a business purpose and separately allows aggregated, de-identified data to be used to analyze and improve the service, but it does not explicitly say whether individual voice or content uploads are used to train Fish Audio's own text-to-speech or voice-cloning models. The terms-of-use page did not surface an explicit AI-training clause in the content that was retrievable, which is a genuine sourcing gap rather than a clean not-published finding. Separately, a different domain with a similar name publishes a shorter privacy policy attributing the service to a different legal entity than fish.audio's stated operator, an unresolved discrepancy that was not used as a source given the requirement to stick to this vendor's own domain.
- Compliance
- Not published
- Retention
- Personal data is retained for as long as necessary to provide the service or fulfill the purpose for which it was collected, with no fixed period published.
- Export
- Partial export
Sources: fish.audio/privacy, fish.audio/terms
As published by the vendor. Verify independently before purchase decisions.
Our Take
Fish Audio's actual story is that it's the TTS infrastructure already running inside HeyGen and Retell in production, available directly as a standalone API and playground. That makes it the logical pick for developers who want voice cloning or speech synthesis without building on top of a consumer-facing product that abstracts away the underlying model. Voice cloning works from a 10-second clip, API pricing starts at $15 per million characters with no minimum commitment, and the genuinely usable free tier covers experimentation. Commercial voice cloning rights require a paid plan, so confirm licensing before deploying.