pantinor/ovos-conformer-mirrors

OpenVoiceOS NVIDIA conformer-transducer mirrors (sherpa-onnx compatible)

Mirrors of the OpenVoiceOS int8 ONNX exports of the NVIDIA NeMo conformer-transducer-large ASR models, re-packaged so that sherpa-onnx (v1.13.x) can load them.

Why a mirror

The upstream exports (e.g. OpenVoiceOS/nvidia-it-conformer-transducer-large-onnx) cannot be loaded by sherpa-onnx:

1. The encoder carries no k2-fsa metadata, so sherpa exits with No model_type in the metadata! (exit 255).
2. The prediction and joint networks ship as a single combined decoder_joint-model.int8.onnx graph, while sherpa's nemo_transducer loader requires separate decoder and joiner sessions.

What was changed

Per language directory (it/, ru/, de/, es/, en/, fr/):

• encoder-model.int8.onnx: byte-identical weights to upstream, with the following metadata_props injected (values from the upstream config.json and the exported graphs):
  • model_type: nemo_transducer
  • vocab_size (upstream vocab size excluding <blk>; sherpa adds 1)
  • subsampling_factor: 4 (from upstream config.json)
  • normalize_type: per_feature
  • pred_rnn_layers: 1, pred_hidden: 640
  • feat_dim: 80
  • plus version, model_author, comment provenance keys
• decoder-model.onnx and joiner-model.onnx: the upstream decoder_joint-model.int8.onnx graph split into two sub-graphs at the internal prediction-network output tensor (exact computation, no weights changed), matching sherpa's decoder/joiner session interface.
• vocab.txt: unchanged from upstream.

Usage

import sherpa_onnx
recognizer = sherpa_onnx.OfflineRecognizer.from_transducer(
    encoder="it/encoder-model.int8.onnx",
    decoder="it/decoder-model.onnx",
    joiner="it/joiner-model.onnx",
    tokens="it/vocab.txt",
    num_threads=2,
    decoding_method="greedy_search",
    model_type="nemo_transducer",
)

Each language was verified by loading with sherpa-onnx 1.13.4 and transcribing a natural-speech sample (correct transcripts on it, ru, de, es, en, fr).

Provenance and license

Source models: NVIDIA NeMo conformer-transducer-large (Apache-2.0), ONNX exports by OpenVoiceOS. Original weights are unchanged; only ONNX graph container metadata and graph splitting were applied. Upstream repos: OpenVoiceOS/nvidia-{it,ru,de,es,en,fr}-conformer-transducer-large-onnx.