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).

Usage

import onnx_asr

model = onnx_asr.load_model("OpenVoiceOS/misterkissi-w2v2-lg-xls-r-300m-zulu-onnx")
print(model.recognize("test.wav"))

Or with the OVOS plugin:

from ovos_stt_plugin_onnx_asr import OnnxASRSTT

stt = OnnxASRSTT(config={"model": "OpenVoiceOS/misterkissi-w2v2-lg-xls-r-300m-zulu-onnx"})

Files

model.onnx / model.onnx.data (when present; small models may be fully inline) -- 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}.

License

CC BY-NC-SA 4.0, inherited from the source model misterkissi/w2v2-lg-xls-r-300m-zulu.