Qwen3 TTS 12Hz 1.7B Base ONNX

--- license: apache-2.0 base_model: Qwen/Qwen3-TTS-12Hz-1.7B-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 1.7B Base — ONNX ONNX export of [Qwen/Qwen3-TTS-12Hz-1.7B-Base](https://huggingface.co/Qwen/Qwen3-TTS-12Hz-1.7B-Base), a multilingual text-to-speech model from Alibaba Cloud's Qwen team with an ECAPA-TDNN speaker encoder for voice cloning from short reference audio. It is not a newly trained model. This is an unofficial community conversion. The 1.7B Base variant is larger than the 0.6B Base and targets higher-quality synthesis where compute budget permits. The smaller variant is in [`tonythethompson/Qwen3-TTS-12Hz-0.6B-Base-ONNX`](https://huggingface.co/tonythethompson/Qwen3-TTS-12Hz-0.6B-Base-ONNX). ## Source | Field | Value | |---|---| | Upstream model | [Qwen/Qwen3-TTS-12Hz-1.7B-Base](https://huggingface.co/Qwen/Qwen3-TTS-12Hz-1.7B-Base) | | Upstream source revision | `fd4b254389122332181a7c3db7f27e918eec64e3` | | Packaging source revision | `ab09194b7e07e645f2165fda2e95ac97c2446b38` | | 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 | Size | Description | |---|---|---| | `speaker_encoder.onnx` + `.data` | ~46 MB | ECAPA-TDNN speaker encoder | | `talker_prefill.onnx` + `.data` | ~5 GB | Talker LM prefill | | `talker_decode.onnx` + `.data` | ~5 GB | Talker LM single-step decode | | `code_predictor.onnx` + `.data` | ~420 MB | Code Predictor | | `tokenizer12hz_encode.onnx` + `.data` | ~182 MB | 12 Hz codec tokenizer (encode) | | `vocoder.onnx` + `.data` | ~437 MB | Vocoder decoder (24 kHz output) | | `embeddings/` | ~1.4 GB | Text/codec embedding tables (.npy) + config | | `tokenizer/` | ~5 MB | BPE tokenizer (vocab.json, merges.txt) | ## Architecture - **Speaker Encoder**: ECAPA-TDNN, encodes ~3 seconds of reference audio into a speaker embedding - **Talker LM**: Autoregressive transformer (larger than 0.6B), generates 12Hz codec tokens - **Code Predictor**: Predicts remaining RVQ codebook groups from the Talker output - **Vocoder**: RVQ dequantize -> decoder -> BigVGAN, 12Hz -> 24kHz audio output ## Intended Use Multilingual text-to-speech for local inference via ONNX Runtime. Synthesizes speech using a ~3-second voice reference clip, enabling voice-consistent output. The 1.7B Base is the higher-quality option; use the 0.6B Base for lower-latency or lower-VRAM deployments. The CustomVoice variant is in [`tonythethompson/Qwen3-TTS-12Hz-1.7B-CustomVoice-ONNX`](https://huggingface.co/tonythethompson/Qwen3-TTS-12Hz-1.7B-CustomVoice-ONNX). ## Runtime Notes - Designed for ONNX Runtime compatible runtimes. - Output sample rate: 24 kHz. - Voice cloning reference: ~3 seconds of reference audio recommended. - Memory: proportionally higher than the 0.6B variant; exact VRAM requirement not documented in this repo. - 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 - No repository-specific audio quality evaluation is documented here. - Voice cloning from very short or noisy reference clips may degrade results. - Multilingual capability inherited from the upstream model; consult the upstream Qwen3-TTS model card for language coverage. - ONNX export fidelity relative to the upstream PyTorch model is not 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 from a short audio reference. - 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/) — inherited from `Qwen/Qwen3-TTS-12Hz-1.7B-Base`. This packaging repo adds no new license terms.