CosyVoice 3

Disclaimer

This specific page uses temporary drafts written and maintained by Generative AI. While verified to be accurate, please cross-verify with source code.

CosyVoice 3 is the current generation of Alibaba/QwenAudio’s multilingual CosyVoice TTS project. TomoriBot wraps the official runtime in servers/tts/cosyvoice3/ and exposes the same POST /synthesize interface used by the other local speech endpoints.

TomoriBot defaults to the official FunAudioLLM/Fun-CosyVoice3-0.5B-2512 checkpoint. It is the current CosyVoice 3 release recommended by upstream, uses the normal unquantized model, and is small enough to run comfortably on a 16 GB NVIDIA GPU while keeping CosyVoice’s low-latency design intact.

What it supports

The current CosyVoice 3 release supports:

• Chinese, English, Japanese, Korean, German, Spanish, French, Italian, and Russian
• 18+ Chinese dialects and accents
• zero-shot voice cloning
• multilingual and cross-lingual voice cloning
• natural-language instructions for language, dialect, emotion, speaking speed, and volume
• fine-grained controls in the upstream runtime, including [breath] and [laughter]
• text-in and audio-out streaming in the upstream runtime

The official CosyVoice 3 examples currently include one important Japanese caveat: Japanese text is shown after conversion to katakana. Japanese is a supported language, but if normal Japanese orthography gives poor pronunciation, converting the synthesis text to katakana is the upstream-recommended workaround.

How TomoriBot maps requests

The wrapper accepts the normal tts-clone fields:

• text
• ref_audio
• ref_text
• instruct
• language

It chooses the current CosyVoice 3 API as follows:

Request	CosyVoice 3 path
Reference audio + transcript	inference_zero_shot
Reference audio without transcript	inference_cross_lingual
instruct or explicit language	inference_instruct2

For the best ordinary cloning quality, provide both the reference audio and its matching transcript. CosyVoice 3’s current instruction API conditions on the reference audio but does not also accept the reference transcript, so requests that use instruct switch to the official inference_instruct2 path.

Style and emotion controls

Register the endpoint with Plain markup. Delivery direction belongs in the endpoint’s global voice_instructions field, not in arbitrary inline bracket tags. This preserves the meaning of the instruction for the whole utterance and avoids treating a script such as [happy] Hello. [sad] Goodbye. as two contradictory global instructions. Native [breath] and [laughter] support is intentionally deferred until TomoriBot can advertise an exact provider-aware tag capability.

The /synthesize instruct field is passed into CosyVoice 3’s instruction conditioning. Examples include sound relieved but still tired, speak as quickly as possible, or speak quietly with restrained excitement.

Streaming

CosyVoice 3 supports bidirectional streaming upstream. The project documents both text-in streaming and audio-out streaming, with first-audio latency as low as roughly 150 ms in its optimized setup.

TomoriBot’s current custom TTS interface expects one complete audio response for a Discord voice message, so this server returns a complete WAV and defaults upstream inference to stream=False. Set COSYVOICE3_UPSTREAM_STREAM=1 only when testing the upstream generator; it does not reduce TomoriBot’s response latency until a streaming voice transport exists.

Hardware

Recommended TomoriBot starting point:

• NVIDIA GPU with 16 GB VRAM
• Python 3.10
• recent NVIDIA driver compatible with CUDA 12
• git
• ffmpeg for TomoriBot voice-sample normalization
• sox and libsox-dev on Linux if upstream audio compatibility issues occur

The model itself is 0.5B parameters and does not need quantization to fit a 16 GB card. The Hugging Face checkpoint download is much larger than the parameter count suggests because it also ships the flow model, speech tokenizers, English text model, and both base and RL LLM weights. Allow roughly 10 GB of disk space for the current model package, plus the Python environment and runtime.

CPU inference is possible through the upstream runtime but is not the recommended path for low-latency Discord voice use.

Installation

Linux / WSL2 (recommended)

From the TomoriBot repository root:

bash servers/tts/cosyvoice3/install-cosyvoice3.sh
servers/tts/cosyvoice3/.venv/bin/python servers/tts/cosyvoice3/server.py

Or start the configured server and TomoriBot together:

bun run launch --cosyvoice3

The installer:

1. checks out the reviewed QwenAudio/CosyVoice commit 074ca6dc9e80a2f424f1f74b48bdd7d3fea531cc recursively into servers/tts/cosyvoice3/CosyVoice/;
2. creates servers/tts/cosyvoice3/.venv;
3. installs the current upstream CosyVoice requirements plus the small wrapper dependency set; and
4. downloads FunAudioLLM/Fun-CosyVoice3-0.5B-2512 at Hugging Face revision 29e01c4e8d000f4bcd70751be16fa94bf3d85a18 into CosyVoice/pretrained_models/Fun-CosyVoice3-0.5B/.

Reruns keep those exact revisions; moving to newer ones means changing both pins in the installer. The installer refuses to reinstall over a runtime checkout that has local changes.

The upstream requirements currently use PyTorch 2.3.1 with the CUDA 12.1 package index, CUDA 12 ONNX Runtime packages on Linux, and TensorRT 10.13 packages on Linux. If you are using hardware that requires a newer PyTorch CUDA build, install a compatible PyTorch build in the server’s venv after the upstream requirements and test it with your driver.

Windows PowerShell

Native Windows is provided as a best-effort path:

.\servers\tts\cosyvoice3\install-cosyvoice3.ps1
.\servers\tts\cosyvoice3\.venv\Scripts\python.exe servers\tts\cosyvoice3\server.py

For NVIDIA GPU use, WSL2 is recommended. The current upstream requirements install GPU ONNX Runtime on Linux but CPU ONNX Runtime on Windows, so WSL2 more closely matches the configuration the CosyVoice project optimizes and tests for low latency.

Register in TomoriBot

Run /providers, choose Add New Custom Endpoint, and configure the speech endpoint:

• Capability: Speech
• API Compatibility: tts-clone
• Endpoint URL: http://127.0.0.1:8017
• Voice Source Mode: Clone
• Script Markup: Plain
• Supports Instruct: Yes

After saving the connection, select it and add a Speech model. A clear model code is Fun-CosyVoice3-0.5B-2512.

Then open /config > Models > Switch Models and activate the CosyVoice 3 speech endpoint.

Assign a persona voice

For normal zero-shot cloning:

1. Prepare a clean 3 to 30 second sample with one speaker and little or no background noise.
2. Open /config under Models > TTS Parameters & Voices and upload the sample.
3. Enter the matching transcript when possible. CosyVoice 3 uses it for the transcript-backed zero-shot path, and it is tokenized as a prompt prefix, so it should describe the audio that is actually used: the first 30 seconds of the clip.
4. Open /config under Persona > Voice and assign that sample to the persona.

CosyVoice’s speech tokenizer works on a 30-second prompt window, and upstream enforces it by failing: its own web UI tells you to keep the prompt audio under 30 seconds, and the tokenizer asserts that limit instead of shortening the audio itself. The local server trims instead, so a longer clip is cut to its first 30 seconds and synthesis proceeds, and the trim is logged on the server’s console.

The trim reads the clip in place, which means the speaker embedding is taken from the same opening 30 seconds as the prompt speech tokens. That pairing is what CosyVoice conditions on, so a long reference loses nothing that the engine would have used. The practical effect is that only the opening 30 seconds of a long upload condition the voice, while the rest is uploaded and stored without being used.

Keeping the assigned sample at 10 to 20 seconds stays inside the window with room to spare, which also keeps the stored transcript aligned with the audio the model reads.

Cross-lingual cloning is supported. The reference speaker can speak a different language from the generated text. If a reference transcript is unavailable, the wrapper uses CosyVoice 3’s dedicated cross-lingual path.

Test with /generate voice-message

Use /generate voice-message to test the active endpoint without waiting for a normal chat turn to choose the voice tool. You can use the persona’s configured sample or upload a one-off sample. When uploading a sample, provide its transcript in the modal when possible.

For expressive delivery, enter a global delivery direction in the modal or let the voice tool send voice_instructions. Keep the spoken script as plain text; arbitrary inline style tags are removed before synthesis rather than being misrepresented as whole-utterance instructions.

Environment variables

Variable	Default	Purpose
COSYVOICE3_MODEL_DIR	CosyVoice/pretrained_models/Fun-CosyVoice3-0.5B	Local checkpoint directory
TOMORI_TTS_HOST	127.0.0.1	Wrapper bind address; see Network access
COSYVOICE3_PORT	8017	Wrapper port
COSYVOICE3_UPSTREAM_STREAM	0	Enable CosyVoice’s internal streaming generator
COSYVOICE3_SPEED	1.0	Global numeric speed multiplier passed to upstream inference
COSYVOICE3_DEFAULT_INSTRUCT	empty	Optional instruction added when a request does not provide one
COSYVOICE3_FP16	0	Ask the official runtime to use its fp16 mode
COSYVOICE3_LOAD_TRT	0	Enable upstream TensorRT loading when properly prepared
COSYVOICE3_LOAD_VLLM	0	Enable upstream vLLM loading when its separate dependencies are installed

The default keeps TensorRT, vLLM, and fp16 off. The normal PyTorch runtime already fits the target 16 GB GPU, is simpler to install, and avoids turning the default path into an optimization-specific setup.

Performance and model variants

Default: base Fun-CosyVoice3-0.5B-2512

This is the recommended default for TomoriBot. It has strong speaker similarity, supports all current CosyVoice 3 cloning and instruction modes, and does not need quantization on a 16 GB GPU.

RL weight

The current checkpoint package also includes llm.rl.pt. Upstream publishes the base and RL results separately. The RL weight improves some content-error metrics, while the base result retains slightly stronger speaker-similarity scores in the published table. Because TomoriBot emphasizes persona voice cloning, the wrapper leaves the normal llm.pt as the default.

The current official loader always reads a file named llm.pt. To experiment with the RL weight without overwriting the default installation, copy the model directory, replace the copy’s llm.pt with llm.rl.pt, and point COSYVOICE3_MODEL_DIR at that copy.

vLLM and TensorRT

CosyVoice 3 also supports optional vLLM and TensorRT paths. Upstream currently documents vLLM 0.11.x+ using the V1 engine and vLLM 0.9.0 as the legacy path. These runtimes have additional version and hardware constraints, so TomoriBot does not install or enable them by default.

Use them only after the ordinary PyTorch server is working. For a Discord voice-message workload, avoiding extra runtime complexity is usually more useful than optimizing an already-small 0.5B model.

License

The current CosyVoice code repository is licensed under Apache License 2.0, and the FunAudioLLM/Fun-CosyVoice3-0.5B-2512 Hugging Face repository is also marked Apache-2.0.

The upstream model card additionally contains a disclaimer saying the displayed content is for academic demonstration and that some examples may come from the internet. An open upstream discussion asks for explicit clarification about how that disclaimer relates to commercial use of the weights. TomoriBot does not redistribute the model. Self-hosters should review the current upstream license and model-card terms for their own deployment, especially before commercial use.