Qwen3-TTS-12Hz-0.6B CustomVoice - Mixed Precision and Mixed Frameworks
This repository packages a mixed-precision, mixed-framework deployment bundle
derived from
Qwen/Qwen3-TTS-12Hz-0.6B-CustomVoice
for multilingual text-to-speech using preset speakers. The bundle uses GGUF,
ONNX, and LiteRT artifacts and supports streaming synthesis at 24 kHz.
✨ Key Highlights
- No intelligibility regression — no intelligibility WER regression was observed against the FP32 Qwen3-TTS CustomVoice implementation.
- Reduced model size — 1.64× smaller than the FP32 Qwen3-TTS CustomVoice implementation.
- Performance uplift — compared with the FP32 Qwen3-TTS CustomVoice implementation:
- RTF is improved by more than 2× on the MacBook Pro® computer with M4 Pro
- RTF is improved by more than 4× on the Android™ Vivo X300 smartphone.
- Faster than real-time — using one Arm® CPU core with SME2 (An RTF lower than 1 is faster than real time):
- RTF is 0.376 on a MacBook Pro® computer with an M4 Pro chip
- RTF is 0.836 on an Android™ Vivo X300 smartphone.
- Time to first chunk — using one Arm® CPU core with SME2
- latency is 0.139 s on a MacBook Pro® computer with an M4 Pro chip
- latency is 0.314 s on an Android™ Vivo X300 smartphone.
- Arm® CPU deployment — optimized for efficient execution on Arm® CPUs.
- Streaming output — returns synthesized speech at 24 kHz in streaming mode.
📦 Model Details
Model Description
Qwen3-TTS-12Hz-0.6B-CustomVoice is a multilingual text-to-speech model that synthesizes speech using a selected preset speaker.
- Developed by: Qwen team, Alibaba Cloud
- Model type: Multilingual text-to-speech with preset-speaker selection
- License: Apache-2.0, inherited from the upstream model and source project
- Base model:
Qwen/Qwen3-TTS-12Hz-0.6B-CustomVoice - Packaged variant:
qwen3_tts_12hz_0.6b_custom_voice - Package form: GGUF, ONNX, and TFLite artifacts
Model Sources
- Base model: https://huggingface.co/Qwen/Qwen3-TTS-12Hz-0.6B-CustomVoice
- Upstream repository: https://github.com/QwenLM/Qwen3-TTS
🚀 Get Started with the Model
🔓 Compute Flow — Early Access
The inference engine for this model package is available through the Compute Flow Early Access Program.
Want to try it?
📩 Contact us at ai-early-access@arm.com to request access.
📊 Quality evaluation
Quality was assessed using WER on a subset of the Seed-TTS corpus. No intelligibility regression was observed against the FP32 Qwen3-TTS CustomVoice implementation.
Settings: English, Aiden speaker, temperature 0.9, top-k 50, and seed 12345. Audio was converted to mono 16 kHz before ASR scoring.
🎯 Performance evaluation
Performance was measured under the following conditions:
- One Arm® CPU core.
- 5 warmups.
- 30 measured runs with a 60-second pause between runs.
- The Android™ Vivo X300 smartphone screen was kept on.
The following methodology and definitions were used:
- The input prompt was: "Hello everyone. I am Jack and I am your personal assistant."
- End-to-end latency is the summed brick execution time for the complete utterance and excludes model setup.
- Average memory is the mean process RSS sampled throughout setup and inference.
- Peak memory is the maximum sampled RSS or high-water mark.
Compared with the FP32 Qwen3-TTS CustomVoice implementation, the optimized model provides the following uplift.
| Metric | MacBook Pro®: FP32 | MacBook Pro®: Optimized | Uplift | Android™ Vivo X300: FP32 | Android™ Vivo X300: Optimized | Uplift |
|---|---|---|---|---|---|---|
| RTF | 0.839 | 0.376 | 2.23× faster | 3.516 | 0.836 | 4.21× faster |
| End-to-end latency | 3.962 s | 1.684 s | 2.35× faster | 16.877 s | 3.878 s | 4.35× faster |
| Time to first chunk | 0.305 s | 0.139 s | 2.19× faster | 1.240 s | 0.314 s | 3.95× faster |
| Average memory | 4619.3 MiB | 2605.9 MiB | 43.6% lower | 5253.7 MiB | 2883.4 MiB | 45.1% lower |
🛠️ Technical Specifications
Objective
Generate streaming 24 kHz speech from text using a selected preset speaker.
Runtime Architecture
| Component role | Framework / format |
|---|---|
| Text tokenization | Hugging Face tokenizer |
| Embeddings | ONNX Runtime / ONNX |
| Talker, code prediction, and speech decoding | llama.cpp / GGUF and LiteRT / TFLite |
Precision and Quantization
- The talker, code predictor, and speech-decoder transformer use Q8_0 quantization.
- Speech-decoder post-processing uses FP16 compute acceleration.
Input Specification
| Input | Description |
|---|---|
| Text prompt | UTF-8 string |
Output Specification
The model returns synthesized 24 kHz audio in streaming mode.
Manifest Configuration
At runtime, users can select the synthesis language and preset speaker and set the number of active code-prediction groups. The packaged defaults are language- and speaker-specific and use 16 active code-prediction groups.
Supported languages are Chinese, English, Japanese, Korean, German, French, Russian, Portuguese, Spanish, and Italian. Automatic language selection is also available.
Repository Contents
qwen3_tts_custom_voice_<language>_manifest.json— language-specific configuration.tokenizer.json— tokenizer model.*.onnx,*.gguf, and*.tflite— model components.assets/— language, speaker, and variant maps.SHA256SUMS— model-bundle checksums for reproducibility.
🗂️ Model and Asset Origin
🔐 Checksums
SHA256SUMS was generated by recursively hashing every regular file in the
model bundle, including files in subdirectories, except the generated root
SHA256SUMS and paths with a dotfile component.
From the model bundle root, verify the checked-out files with:
shasum -a 256 -c SHA256SUMS