Overview

CosyVoice combines speech synthesis with voice cloning and tools for training and deployment. The repository is licensed under Apache License 2.0, and commercial use is permitted. The maintainers describe its intended use as academic and technical demonstration, so teams should weigh that stated purpose alongside the commercial-use permission.

Installation uses Conda and Python 3.10, with pretrained models downloadable through ModelScope or Hugging Face. Docker deployment, FastAPI and gRPC servers, and NVIDIA Triton and TensorRT-LLM acceleration give infrastructure teams several routes to integration. Those options require teams to handle installation and deployment themselves rather than relying on a turnkey hosted service.

Key features

Multilingual synthesis and cloning: Fun-CosyVoice 3.0 is designed for zero-shot multilingual synthesis. Version 3.0 supports nine languages and more than 18 Chinese dialects or accents, including multilingual and cross-lingual zero-shot voice cloning. This breadth is useful for teams creating speech across language variants, though the facts do not establish parity across every language or accent.
Fine-grained speech direction: Users can provide instructions for language, dialect, emotion, speed, and volume. Pronunciation inpainting accepts Chinese Pinyin and English CMU phonemes, offering a route to correct specific pronunciations without changing the broader workflow.
Text handling and streaming: CosyVoice can read numbers, special symbols, and varied text formats without a traditional frontend module. Text-in, audio-out streaming has latency stated as low as 150 ms, a useful target for responsive applications rather than a guarantee for every deployment.
Deployment interfaces: FastAPI and gRPC servers and clients support API integration, while Docker, Triton, and TensorRT-LLM offer deployment and acceleration options. WAV is the stated export format, so teams needing other output formats should account for that constraint.

Pricing

Open-source software: 0.00 USD per free. The free package includes downloadable source code and models under Apache-2.0, with self-managed installation and deployment. There are no paid plans in the stated offering. This avoids a subscription charge, but teams take on the infrastructure and operating work themselves.

The repository documents Python 3.10. The optional ttsfrd normalization wheel is specified for Linux x86_64, and CosyVoice 2 and 3 support vLLM 0.11.x or newer and vLLM 0.9.0; versions between those releases are untested. These constraints matter for teams integrating with an existing runtime.

Platforms

CosyVoice supports API use, Linux, and self-hosted deployment. The documented setup uses Conda and Python 3.10, and offers Docker deployment with FastAPI or gRPC, plus NVIDIA Triton and TensorRT-LLM options. Maintainer support is through GitHub Issues and an official Dingding chat group.

Who it's for

CosyVoice fits developers and teams that can operate their own speech stack and want multilingual synthesis, zero-shot voice cloning, and configurable deployment. The language and dialect coverage and API interfaces make it relevant to applications that need speech generation integrated into existing infrastructure. It is a weaker fit for buyers seeking a managed web or mobile product, or teams unable to accommodate the documented Python and vLLM version requirements.

Pros and cons

Pros: Nine languages and more than 18 Chinese dialects or accents give multilingual projects substantial stated coverage.
Pros: Voice cloning, pronunciation inpainting, and controls for language, dialect, emotion, speed, and volume support more tailored output.
Pros: Free source and models, Apache-2.0 licensing, and API deployment options let teams self-manage integration without a listed subscription plan.
Cons: Installation and deployment are self-managed, which shifts setup and operations to the adopting team.
Cons: Python 3.10 and the stated vLLM compatibility ranges can complicate adoption in stacks pinned to other versions.
Cons: WAV is the stated export format, and the maintainers frame the repository for academic use and technical demonstration.

Alternatives

Qwen3-TTS is another free, self-hosted, Apache-2.0 option; its open-source models include 0.6B and 1.7B models and support 10 languages. Choose it when that model selection or language count better matches the project.

Fish Audio offers a free tier with 8,000 monthly credits, up to seven minutes of generation, and up to 500 characters per generation. Its API, web, desktop, and self-hosted platform coverage may suit teams looking for a broader set of access routes and a defined free quota.

Fish Speech is a free option with API, web, Linux, and self-hosted platforms; its free tier includes 8,000 monthly credits, up to seven minutes of generation, and three public voice slots. It may fit teams who prefer a usage-capped free tier.

Sprag is a paid alternative with stock-voice text-to-speech priced at 0.70 USD per month, billed at $0.70 per hour of speech generated and also listed at $19.44 per 1M audio tokens. Choose it when that metered stock-voice offering fits better than self-managed deployment.

Kits AI has a free plan with 15 conversion minutes, zero voice slots, and zero download minutes. Its stated inclusion of voice designer/blender and generative vocals may make it relevant when those capabilities matter more than downloadable voice slots.

Voicebox is another alternative.

VoiceStudio is another alternative.

GPT-SoVITS is another free alternative with API and self-hosted platform support.

For more options, browse Voice Cloning Software or Text-to-Speech Software.

Verdict

Choose CosyVoice if your team can manage its own deployment and wants multilingual synthesis and voice cloning with API integration and no listed subscription cost. Look elsewhere if you need a managed service, a broader stated export-format choice, or a project purpose better aligned with a production-focused offering.

	CosyVoice	T-Bank VoiceKit
Free plan	Yes	No
Free trial	No	No
Paid from	Free	—
Open source	No	No
Platforms	api, Linux, self-hosted	api
Free plan	Yes	—
Commercial use	Yes	—
Voice cloning	Yes	—
API access	Yes	Yes
Export formats	WAV	LINEAR16, ALAW, RAW_OPUS
Platforms	linux, api, self_hosted	api

CosyVoice vs T-Bank VoiceKit

CosyVoice

T-Bank VoiceKit