OpenVoiceOS/wav2vec2-large-sami-22k-onnx

wav2vec2-large-sami-22k (ONNX)

ONNX export of 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 (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

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.