ONNX export of hexgrad/Kokoro-82M, an open-weight English text-to-speech model, in a Q8/FP16 mixed-precision format. It is not a newly trained model. This is an unofficial community conversion; hexgrad is the original model author.
--- license: apache-2.0 base_model: hexgrad/Kokoro-82M base_model_relation: quantized pipeline_tag: text-to-speech library_name: onnxruntime language: - en tags: - onnx - onnxruntime - kokoro - tts - text-to-speech - q8f16 - quantized --- # Kokoro-82M — Q8/FP16 ONNX ONNX export of [hexgrad/Kokoro-82M](https://huggingface.co/hexgrad/Kokoro-82M), an open-weight English text-to-speech model, in a Q8/FP16 mixed-precision format. It is not a newly trained model. This is an unofficial community conversion; hexgrad is the original model author. "Q8F16" refers to 8-bit weights with FP16 activations (or FP16 weights with Q8 compute paths, depending on the export configuration). Quantization follows the ONNX Community `model_q8f16.onnx` export. This variant targets CPU or lower-VRAM GPU execution where the FP16 model is too large or too slow. ## Source | Field | Value | |---|---| | Upstream model | [onnx-community/Kokoro-82M-v1.0-ONNX](https://huggingface.co/onnx-community/Kokoro-82M-v1.0-ONNX) | | Upstream source revision | `1939ad2a8e416c0acfeecc08a694d14ef25f2231` | | Export tool/script | Mirrored from onnx-community (no re-export) | | Quantization recipe | Pre-quantized Q8F16 ONNX mirrored from `onnx/model_q8f16.onnx` | ## Files | File | Size | Description | |---|---|---| | `onnx/model_q8f16.onnx` | ~82 MB | Q8/FP16 ONNX model (self-contained) | ## Intended Use An English text-to-speech model packaged for local inference via ONNX Runtime. The Q8/FP16 variant targets CPU or constrained-GPU execution where the FP16 model exceeds the available memory or compute budget. The FP16 variant is in [`tonythethompson/kokoro-82m-fp16-onnx`](https://huggingface.co/tonythethompson/kokoro-82m-fp16-onnx). ## Runtime Notes - Designed for ONNX Runtime compatible runtimes. - Target: CPU or GPU with limited VRAM. - Sample rate: 24 kHz (inherited from Kokoro-82M). - Mixed quantization: execution-provider behavior depends on the ONNX Runtime version and hardware. 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 - English-only TTS. - No parity evaluation against the original PyTorch or FP16 ONNX model is documented here. - Audio quality degradation relative to FP16 is possible; validate on your target use case. ## Safety and Responsible Use Kokoro-82M produces realistic synthetic speech (from preset voice styles, not arbitrary speaker cloning). - 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 in their jurisdiction. ## License [Apache 2.0](https://choosealicense.com/licenses/apache-2.0/) — inherited from `hexgrad/Kokoro-82M`. This packaging repo adds no new license terms.
ONNX export of hexgrad/Kokoro-82M, an open-weight English text-to-speech model, in a Q8/FP16 mixed-precision format. It is not a newly trained model. This is an unofficial community conversion; hexgrad is the original model author.
