Fish Speech

Overview

Fish Speech is an open-source multilingual text-to-speech system for generating expressive speech, with both self-hosted and service-oriented ways to run it. It suits developers and teams who want control over voice generation and can meet its hardware and licensing requirements. Its strongest case is flexible speech generation; its clearest caveat is that the model license does not permit commercial use without a separate written license.

Key features

S2 combines broad language coverage with controls useful for more than straightforward narration. Its natural-language tags allow fine-grained adjustment of emotion and prosody, while the repository describes S2 Pro as supporting more than 80 languages without phonemes or language-specific preprocessing. That breadth is attractive for multilingual projects, though it does not remove the need to plan for the listed hardware requirements.

Voice cloning uses short reference samples, typically 10–30 seconds, without additional fine-tuning. S2 can also take multiple speaker references and speaker ID tokens to generate dialogue with multiple speakers in one pass. These capabilities make it a more flexible fit for character dialogue or voiced scenes than a simple single-voice workflow.

For serving, the documentation describes SGLang optimizations including continuous batching, paged KV cache, CUDA graphs, and prefix caching. These are relevant to teams building generation pipelines, rather than a reason on their own for an individual user to choose the system.

Pricing

The model is open source and free for research and non-commercial use under the Fish Audio Research License. Commercial use requires a separate written license from Fish Audio; the free model license does not grant those rights. The service plans below have listed prices, but those prices do not change the model license’s commercial-use restriction.

The Free Tier is a bounded way to try generation, but its short monthly allowance, 500-character cap, and public-only voice slots constrain sustained or private work. Plus adds priority speed and one professional voice slot, making it a modest step up for an individual. Pro is the first listed tier with team seats; Max expands both seats and monthly capacity for larger workloads. Enterprise is the fit for organizations needing on-premise deployment or zero data retention, but it uses custom pricing.

Platforms

Fish Speech supports API, Linux, self-hosted, and web use. The installation guide documents command-line, WebUI, server, and Docker setups, including CPU installation and GPU deployment. However, its stated inference requirement is 24 GB of GPU memory, with Linux or WSL as system requirements. That makes local deployment a poor match for users without suitable GPU resources or a compatible system.

Enterprise support includes 24/7 production support, a technical account manager, and a 99% uptime SLA. Fish Audio says its SOC 2 Type II audit is underway and offers zero data retention, HIPAA-aligned configurations, and self-hosted deployments for enterprise workloads. Its enterprise integrations include real-time pipelines, WebRTC infrastructure, workflow automation, Vapi, Twilio, and Retell.

Who it's for

Fish Speech is best for developers, researchers, and production teams that need multilingual speech generation, voice cloning, or multi-speaker output and can manage the deployment. It is especially compelling when natural-language emotion controls and self-hosting matter. Casual users seeking a lightweight local install should look elsewhere if they cannot meet the 24 GB GPU-memory requirement; commercial users should secure written licensing before relying on it.

Pros and cons

• Pro: Fine-grained emotion and prosody tags provide expressive control beyond basic text-to-speech.
• Pro: Short-sample cloning and multi-speaker generation support varied voice and dialogue workflows without additional fine-tuning.
• Pro: Command-line, WebUI, server, and Docker options give technical users several deployment routes.
• Con: The 24 GB GPU-memory requirement and Linux or WSL requirement make self-hosting demanding for many personal devices.
• Con: Commercial use requires a separate written license, limiting what the free research and non-commercial license allows.
• Con: The free service tier caps monthly generation at 7 minutes and limits each generation to 500 characters.

Alternatives

For a broader directory of voice-cloning options, browse Voice Cloning Software; for game-focused voice generation choices, see Game Voice Generators.

• Fish Audio is the related freemium option, with web and desktop-platform support in addition to API, Linux, and self-hosting.
• Voicebox is another free-plan option spanning mobile, desktop, API, and self-hosted platforms.
• VoiceStudio is a free alternative with API, web, desktop, and self-hosted platform support.
• VoxCPM is a free alternative with no paid tiers or usage limits stated.
• Kits AI is a freemium alternative whose free plan includes 15 conversion minutes, but no voice slots or download minutes.
• Qwen3-TTS offers open-source models under Apache-2.0, including 0.6B and 1.7B models supporting 10 languages.
• Sprag is a paid API, self-hosted, and web option with stock-voice text-to-speech priced at $0.70 a month for Chorus.
• Chatterbox TTS offers free open-source models under the MIT license for research and development, across Linux, macOS, and self-hosted use.

Verdict

Choose Fish Speech if you need expressive, multilingual speech generation with cloning and multi-speaker controls, and have the technical capacity to deploy it. Its flexible workflows make it a strong option for developers and research teams; the hardware burden and separate commercial licensing make it a less straightforward choice for casual users or businesses seeking immediate commercial rights.