--- license: apache-2.0 base_model: Qwen/Qwen3-TTS-12Hz-0.6B-Base base_model_relation: quantized pipeline_tag: text-to-speech library_name: onnxruntime language: - multilingual tags: - onnx - onnxruntime - tts - qwen3-tts - text-to-speech - voice-cloning - ecapa-tdnn --- # Qwen3-TTS 12Hz 0.6B Base — ONNX ONNX export of [Qwen/Qwen3-TTS-12Hz-0.6B-Base](https://huggingface.co/Qwen/Qwen3-TTS-12Hz-0.6B-Base) for local inference with ONNX Runtime. Includes an ECAPA-TDNN speaker encoder for voice cloning from ~3 seconds of reference audio. This is an unofficial community mirror of the ONNX export; it is not a newly trained model. The Qwen team (Alibaba Cloud) is the original author. ## Source | Field | Value | |---|---| | Upstream model | [Qwen/Qwen3-TTS-12Hz-0.6B-Base](https://huggingface.co/Qwen/Qwen3-TTS-12Hz-0.6B-Base) | | Upstream source revision | `5d83992436eae1d760afd27aff78a71d676296fc` | | Packaging source revision | `17a2fccf89a5391005f9ff163b07e13f7814dddf` | | Export tool/script | ONNX export from upstream Qwen3-TTS PyTorch weights (community packaging) | | Quantization recipe | See `onnx/` filenames for FP32/FP16/quant variants shipped in this repo | ## Files | File | Description | Size | |---|---|---:| | `speaker_encoder.onnx` + `.data` | ECAPA-TDNN speaker encoder | ~34 MB | | `talker_prefill.onnx` + `.data` | Talker LM prefill (28 layers) | ~1.7 GB | | `talker_decode.onnx` + `.data` | Talker LM single-step decode | ~1.7 GB | | `code_predictor.onnx` | Code Predictor (5 layers, 15 groups) | ~440 MB | | `vocoder.onnx` | Vocoder decoder (24kHz output) | ~2.7 MB | | `embeddings/` | Text/codec embeddings as .npy + config | ~1.4 GB | | `tokenizer/` | BPE tokenizer (vocab.json, merges.txt) | ~4 MB | ## Architecture - **Speaker Encoder**: ECAPA-TDNN, 128 mel bins input, 1024-dim speaker embedding output - **Talker**: 28 transformer layers, 16 attn heads, 8 KV heads, hidden=1024 - **Code Predictor**: 5 layers, generates codebook groups 1-15 from Talker output - **Vocoder**: RVQ dequantize -> transformer -> BigVGAN decoder, 12Hz codec -> 24kHz audio ## Intended Use Multilingual text-to-speech for local inference via ONNX Runtime. The Base variant synthesizes speech conditioned on a speaker embedding extracted from a short reference clip, enabling voice-consistent synthesis. The 1.7B Base variant is in [`tonythethompson/Qwen3-TTS-12Hz-1.7B-Base-ONNX`](https://huggingface.co/tonythethompson/Qwen3-TTS-12Hz-1.7B-Base-ONNX); predefined-speaker variants are in the CustomVoice repos. ### Standalone usage (external project) The snippet below uses the external [ElBruno/QwenTTS](https://github.com/elbruno/ElBruno.QwenTTS) C# wrapper and references that project's ONNX repo, not this one. It is included only as a reference for standalone C# use. ```bash dotnet add package ElBruno.QwenTTS.VoiceCloning ``` ```csharp using ElBruno.QwenTTS.VoiceCloning.Pipeline; var cloner = await VoiceClonePipeline.CreateAsync(); await cloner.SynthesizeAsync("Hello world!", "reference.wav", "output.wav", "english"); ``` ## Runtime Notes - Designed for ONNX Runtime compatible runtimes. - Output sample rate: 24 kHz. - Voice cloning reference: ~3 seconds of reference audio recommended. - Validate on the target execution provider before production use. ## Precision and Packaging Export tooling, precision, and quantization are recorded in the **Source** table above. This packaging mirror does not publish independent parity benchmarks; validate on your target execution provider before production use. ## Limitations - Voice cloning quality depends on reference audio quality; noisy or very short clips may degrade results. - Multilingual capability: verify the upstream model card for supported languages and quality by language. - No repository-specific audio quality evaluation is documented here. ## Safety and Responsible Use Qwen3-TTS is a voice synthesis and voice cloning model capable of producing realistic speech closely matching a target speaker. - Do not use to impersonate real individuals without their explicit consent. - Do not generate synthetic speech intended to deceive listeners about a speaker's identity. - Disclose AI-generated audio where listeners would reasonably expect a human voice. - Users are responsible for compliance with applicable laws governing synthetic media and voice cloning in their jurisdiction. ## License [Apache 2.0](https://choosealicense.com/licenses/apache-2.0/) — same as `Qwen/Qwen3-TTS-12Hz-0.6B-Base`. This packaging repo adds no new license terms.
