asr-uz-fastconformer-large-onnx

ASR model exported to ONNX for the onnx-asr library and the ovos-stt-plugin-onnx-asr OpenVoiceOS STT plugin. Ships both the RNNT head (default) and the CTC head.

Language: Uzbek. Converted from uzinfocom-edu-ai/asr-uz-fastconformer-large, a FastConformer Hybrid Transducer+CTC model (EncDecHybridRNNTCTCBPEModel, ~120M parameters) fine-tuned on ~701 hours of curated Uzbek speech.

Reported WER (RNNT, source model card): 16.01%.

Usage

onnx-asr

import onnx_asr
model = onnx_asr.load_model("OpenVoiceOS/asr-uz-fastconformer-large-onnx")
print(model.recognize("audio.wav"))

ovos-stt-plugin-onnx-asr

stt:
  module: ovos-stt-plugin-onnx-asr
  ovos-stt-plugin-onnx-asr:
    model: OpenVoiceOS/asr-uz-fastconformer-large-onnx
    lang: uz

Credits

Original model by uzinfocom-edu-ai, released under the Apache 2.0 license: uzinfocom-edu-ai/asr-uz-fastconformer-large.