Italian speech-to-text model. ONNX export of nvidia/stt_it_fastconformer_hybrid_large_pc — an NVIDIA NeMo FastConformer-Hybrid CTC model — for onnx-asr. Runs offline with ONNX Runtime; PyTorch and NeMo are not required.

Part of the OpenVoiceOS STT/ASR ONNX collection.

Files

File	Purpose
`model.onnx`	Encoder + CTC head, fp32
`vocab.txt`	Token vocabulary (`<token> <id>` per line, `▁` = space, `<blk>` = CTC blank)
`config.json`	onnx-asr metadata: `model_type: nemo-conformer-ctc`, `features_size: 80`, `subsampling_factor: 8`

There is no int8 variant: these architectures are convolution-dominated, and dynamic quantization produces ConvInteger nodes that ONNX Runtime cannot execute on CPU. int8 requires static QDQ quantization with calibration data.

Usage

With onnx-asr (pip install onnx-asr[cpu,hub]):

python

import onnx_asr

model = onnx_asr.load_model("OpenVoiceOS/stt_it_fastconformer_hybrid_large_pc_onnx")
print(model.recognize("speech.wav"))  # 16 kHz PCM wav

With OpenVoiceOS, through ovos-stt-plugin-onnx-asr (mycroft.conf):

json

{
  "stt": {
    "module": "ovos-stt-plugin-onnx-asr",
    "ovos-stt-plugin-onnx-asr": {
      "model": "OpenVoiceOS/stt_it_fastconformer_hybrid_large_pc_onnx"
    }
  }
}

Export and verification

Exported from the original checkpoint with NeMo's model.export() (CTC decoder head of the hybrid model; see the conversion guide). The subsampling_factor was measured empirically on the exported graph, and the export was verified differentially: the ONNX model and the original NeMo checkpoint produce identical transcriptions on a reference clip.

Accuracy, training data and limitations

See the source model card for benchmark results, training corpora and known limitations. This repo changes the runtime, not the weights.

Related projects

onnx-asr — ASR inference with ONNX Runtime
ovos-stt-plugin-onnx-asr — OpenVoiceOS STT plugin
NVIDIA NeMo — framework the source model was trained with