Wav2vec2 Large Sami 22k Onnx

--- license: apache-2.0 language: - smi tags: - onnx - onnx-asr - wav2vec2 - automatic-speech-recognition base_model: - GetmanY1/wav2vec2-large-sami-22k-finetuned --- # wav2vec2-large-sami-22k (ONNX) ONNX export of [GetmanY1/wav2vec2-large-sami-22k-finetuned](https://huggingface.co/GetmanY1/wav2vec2-large-sami-22k-finetuned), a North Sámi wav2vec2 CTC ASR model (Getman & Grósz), pretrained on 22.4k hours and fine-tuned on 20h of North Sámi Parliament session audio, for use with [onnx-asr](https://github.com/istupakov/onnx-asr) (`wav2vec2-ctc` model type). Per-utterance zero-mean/unit-variance normalization is baked into the ONNX graph, masked by `input_lengths` for correct behavior with padded/batched input, so the model works with onnx-asr's plain `identity` preprocessor (raw 16kHz waveform in). > Note: the source model reports WER 33.32 / CER 12.76 on an out-of-domain 1h > test set — treat as a usable-but-imperfect low-resource model, not > production-grade broadcast-quality ASR. ## Usage ```py import onnx_asr model = onnx_asr.load_model("OpenVoiceOS/wav2vec2-large-sami-22k-onnx") print(model.recognize("test.wav")) ``` ## Files * `model.onnx` / `model.onnx.data` — fp32 ONNX graph (inputs: `input_values` (batch, samples) float32, `input_lengths` (batch,) int64; output: `logprobs` (batch, frames, vocab) float32 log-softmax). * `vocab.txt` — CTC vocabulary in onnx-asr's `token id` format (word-delimiter -> `▁`, pad token -> `<blk>`). * `config.json` — `{"model_type": "wav2vec2-ctc", "subsampling_factor": 320}`. No int8 quantized variant is included yet -- `onnxruntime.quantization` does not currently support the `torch.onnx` dynamo-exported graph for this architecture. ## License Apache License 2.0, inherited from the base model.