Qwen3 TTS#

Qwen3-TTS-12Hz-Base is a discrete multi-codebook text-to-speech model from the Qwen team. It performs fast voice cloning from a short reference clip, supports 10 languages, and streams 24 kHz speech with low latency. The 12Hz in the name refers to the codec frame rate (12 acoustic frames per second), not the playback sample rate. SGLang-Omni serves two checkpoints — 0.6B and 1.7B — through the same preprocessing → tts_engine → vocoder pipeline and the OpenAI-compatible /v1/audio/speech endpoint.

Prerequisites#

Install sglang-omni by following Installation.

Qwen3-TTS Base uses the upstream qwen-tts package, which currently pins Transformers 4.57.3. Install it only in environments that serve Qwen3-TTS:

uv pip install transformers==4.57.3 accelerate==1.12.0 sox einops
uv pip install --no-deps qwen-tts==0.1.1

Do not add --upgrade here. It pulls a newer torch/numpy/CUDA stack and breaks inference (mismatched cuDNN, numba requires NumPy ≤ 2.3). Pin only what is listed above so the image’s existing torch build is left untouched.

Download a checkpoint (both repositories are public, no token required):

hf download Qwen/Qwen3-TTS-12Hz-0.6B-Base
hf download Qwen/Qwen3-TTS-12Hz-1.7B-Base

Server Configuration#

The pipeline is preprocessing → tts_engine → vocoder.

# 0.6B
sgl-omni serve \
  --model-path Qwen/Qwen3-TTS-12Hz-0.6B-Base \
  --config examples/configs/qwen3_tts_0_6b.yaml \
  --port 8000

# 1.7B
sgl-omni serve \
  --model-path Qwen/Qwen3-TTS-12Hz-1.7B-Base \
  --config examples/configs/qwen3_tts_1_7b.yaml \
  --port 8000

Synthesizing Speech#

Zero-shot#

Qwen3-TTS does not support zero-shot synthesis.

Voice Cloning#

The references field accepts audio_path (a local path or HTTP URL) and text (the transcript of that clip). Supplying the transcript enables in-context-learning (ICL) mode and materially improves cloning quality; omitting it falls back to speaker-embedding (x-vector) mode.

curl -X POST http://localhost:8000/v1/audio/speech \
  -H "Content-Type: application/json" \
  -d '{
    "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",
      "text": "We asked over twenty different people, and they all said it was his."
    }]
  }' \
  --output output.wav

ref_audio and ref_text are accepted as shorthand for references[0].audio_path and references[0].text.

Python#

import requests

resp = requests.post(
    "http://localhost:8000/v1/audio/speech",
    json={
        "input": "Get the trust fund to the bank early.",
        "references": [{
            "audio_path": "https://huggingface.co/datasets/zhaochenyang20/seed-tts-eval-mini/resolve/main/en/prompt-wavs/common_voice_en_10119832.wav",
            "text": "We asked over twenty different people, and they all said it was his.",
        }],
    },
)
resp.raise_for_status()
with open("output.wav", "wb") as f:
    f.write(resp.content)

Language Hint#

language biases the model toward a target language. It defaults to auto (let the model detect). Supported languages are Chinese, English, Japanese, Korean, German, French, Russian, Portuguese, Spanish, and Italian.

curl -X POST http://localhost:8000/v1/audio/speech \
  -H "Content-Type: application/json" \
  -d '{
    "input": "今天天气不错，就该出去晒晒太阳。",
    "references": [{
      "audio_path": "https://huggingface.co/datasets/zhaochenyang20/seed-tts-eval-mini/resolve/main/en/prompt-wavs/common_voice_en_10119832.wav",
      "text": "We asked over twenty different people, and they all said it was his."
    }],
    "language": "Chinese"
  }' \
  --output output.wav

Streaming#

Set "stream": true to receive audio chunks in real time over Server-Sent Events (SSE):

curl -N -X POST http://localhost:8000/v1/audio/speech \
  -H "Content-Type: application/json" \
  -d '{
    "input": "Get the trust fund to the bank early.",
    "references": [{
      "audio_path": "https://huggingface.co/datasets/zhaochenyang20/seed-tts-eval-mini/resolve/main/en/prompt-wavs/common_voice_en_10119832.wav",
      "text": "We asked over twenty different people, and they all said it was his."
    }],
    "stream": true
  }'

Each event carries a base64-encoded audio chunk; the stream ends with data: [DONE]. See the Higgs TTS cookbook for a full Python SSE consumer.

Generation Parameters#

Parameter	Default	Notes
`input`	(required)	Text to synthesize
`references`	`null`	Reference clip for cloning; each item has `audio_path` and `text`
`ref_audio` / `ref_text`	`null`	Shorthand for `references[0].audio_path` / `references[0].text`
`language`	`auto`	Target-language hint (see list above)
`temperature`	`0.9`	Sampling temperature
`top_p`	`1.0`	Top-p sampling
`top_k`	`50`	Top-k sampling
`repetition_penalty`	`1.05`	Repetition penalty
`max_new_tokens`	`2048`	Maximum number of generated codec tokens
`seed`	`null`	Random seed for reproducibility
`stream`	`false`	Stream audio chunks over SSE

Model Variants#

Checkpoint	Parameters	Config
`Qwen/Qwen3-TTS-12Hz-0.6B-Base`	0.6B	`examples/configs/qwen3_tts_0_6b.yaml`
`Qwen/Qwen3-TTS-12Hz-1.7B-Base`	1.7B	`examples/configs/qwen3_tts_1_7b.yaml`

Both expose an identical request API. The 1.7B model has higher capacity (typically better quality) at a larger memory and latency cost; the 0.6B model is lighter and faster.

Benchmark Results#

Qwen3-TTS-12Hz-0.6B-Base on Seed-TTS EN (1088 utterances, reference voice cloning from each prompt), concurrency 16, WER scored with HF Whisper-large-v3. Hardware: 1× H200 SXM.

Metric	Value
WER (corpus, excl. runaway outliers)	1.07%
WER (per-sample median / p95)	0.00% / 9.09%
WER (corpus micro-avg, raw)	18.29%
Runaway samples (>50% WER)	2 / 1088 (0.2%)
Latency mean / median (s)	6.61 / 6.24
RTF mean / median	1.51 / 1.48
Output throughput (tok/s)	115.4
Completed / failed requests	1088 / 0

Typical output is clean (0.00% median WER, 9.09% p95). Two utterances (0.2%) ran away into a repetition loop and generated ~164 s of looping audio up to max_new_tokens, which alone lifts the raw micro-average to 18.29%; excluding those, corpus WER is 1.07%. RTF > 1 reflects the 0.6B codec pipeline at concurrency 16, not single-stream latency. The 1.7B checkpoint trades latency for quality.

Known Limitations#

Reference audio recommended. As a cloning model, Qwen3-TTS Base produces robotic speech without a reference clip.
Transcript improves cloning. Providing text in references (ICL mode) yields better speaker similarity than speaker-embedding-only (x-vector) mode.
Language detection. language: auto may misdetect for short or code-switched inputs; set language explicitly when you know the target language.
Rare runaway generation. Roughly 0.2% of utterances (observed on the 0.6B checkpoint) can fall into a repetition loop and keep generating up to max_new_tokens. Raising repetition_penalty (default 1.05) or lowering max_new_tokens mitigates it; the 1.7B checkpoint is less prone.

Qwen3 TTS

Contents