Zonos3#

Zonos3-Mini is Zyphra’s text-to-speech model. Generate 44.1 kHz speech with a default voice or clone a voice from reference audio through the OpenAI-compatible /v1/audio/speech API.

Quick start#

Installation requirements#

Follow the CUDA installation guide, using a checkout that includes Zonos3 support. No additional model-specific dependencies are required. Your Hugging Face account must have access to Zyphra-staging/Zonos3-Mini-Staging-Release; authenticate the server environment before launching.

Start the server and check readiness#

sgl-omni serve \
  --model-path Zyphra-staging/Zonos3-Mini-Staging-Release \
  --port 8000

In another terminal, check readiness after the weights load:

curl -fS http://localhost:8000/health

Continue when this returns HTTP 200 with "status": "healthy". During startup it may return HTTP 503 or refuse the connection; check the server logs and retry once initialization finishes.

Generate output.wav with the default voice#

No reference audio is required. Send text to the running server:

curl -fS -X POST http://localhost:8000/v1/audio/speech \
  -H "Content-Type: application/json" \
  -d '{
    "input": "SGLang-Omni is a great project!"
  }' \
  --output output.wav

The response is a complete WAV file containing 44.1 kHz mono speech. Open output.wav in an audio player.

Voice cloning#

Add references to the same request to use a reference voice. No transcript is needed. Zonos3 accepts at most one reference clip, at least 0.5 seconds long. By default, clips longer than 60 seconds are truncated to their first 60 seconds.

{
  "input": "SGLang-Omni is a great project!",
  "references": [{
    "audio_path": "https://huggingface.co/datasets/zhaochenyang20/seed-tts-eval-mini/resolve/main/en/prompt-wavs/common_voice_en_10119832.wav"
  }]
}

For reference formats and the ref_audio shorthand, see Voice Cloning; for server-visible local paths and remote URL access, see media access rules.

Streaming#

Add "stream": true and "response_format": "pcm" to either request above to receive audio while generation is still running.

Use the shared streaming examples or Python streaming client with the request fields above. Zonos3 outputs 44.1 kHz, signed 16-bit little-endian mono PCM. The Python client reads the sample rate from the response headers and supports this format.

Streaming PCM can differ from non-streaming WAV even with the same seed. The default first chunk targets 0.8 s of audio; this is playback duration, not the wait for the first chunk.

Common options#

Add these fields to the JSON request as needed:

Parameter	Default	Usage
`max_new_tokens`	`2048`	Positive generation budget in delay-pattern rows, not text tokens or seconds.
`temperature`	`0.75`	Nonnegative; `0` selects greedy decoding.
`top_k`	`80`	Top-k sampling; integer from `1` to `4098`.
`top_p`	`0.69`	Applied after top-k; must be in `(0, 1]`.
`seed`	unset	Integer sampling seed; omitted requests draw a random seed.
`repetition_penalty`	`2.4`	Must be at least `1.0`; `1.0` disables the penalty.

Text, reference audio, and the generation budget must fit the server’s default 8192-row context window. The generation budget is an upper limit, not a requested audio duration. Shared fields are documented in Request Parameters.

Next steps#

Use the shared Python requests, OpenAI Python SDK, and Python streaming client examples with the checkpoint ID above. Remove the Qwen-specific language field; omit voice in HTTP requests or use voice="default" in the SDK.

Benchmark results#

SeedTTS EN voice cloning (1088 samples), one H100 per server, concurrency 16, default server and sampling settings, no request seed, scored with Qwen3-ASR-1.7B. Values are means of 30 passes: six per warm server on five GPUs. SIM is WavLM-large speaker similarity to the reference, in percentage points.

Mode	WER (corpus)	WER, samples ≤50%	Samples above 50% WER	SIM	RTF mean	Throughput (qps)	First audio (s)
Non-streaming	3.03%	1.26%	21.1	50.8	0.060	55.67	-
Streaming	3.13%	1.25%	22.2	50.8	0.070	48.03	0.14

Corpus WER includes all scored samples. The filtered WER excludes individual samples above 50% WER and is diagnostic only; 21.1 and 22.2 are mean counts across passes. First audio measures the wait for the first audio payload, not its playback duration. See the TTS benchmark guide for RTF, throughput, and other metric definitions.

Throughput by concurrency on the same H100 setup: one pass per level, except concurrency 16, which repeats the 30-pass means above.

Concurrency	1	4	16	32	64
Non-streaming (qps)	9.05	25.61	55.67	70.32	74.35
Streaming (qps)	8.54	23.22	48.03	57.97	61.66
Streaming first audio (s)	0.057	0.078	0.140	0.221	0.264

These original measurements did not record an immutable checkpoint revision and have not been rerun for the staging checkpoint linked above.

Zonos3

Contents