sherpa-onnx/python-api-examples/orukeet.md

Orukeet with the offline WebSocket server

Orukeet is a 25-language fine-tune of Parakeet TDT 0.6B v3. Its ONNX export uses sherpa-onnx's existing NeMo transducer implementation. The fitted Gabor filters are stored as convolution weights; inference requires no custom operators.

Download

The following commands download the release's manifest and INT8 model archive from an immutable Hugging Face revision. The manifest supplies the archive's size, checksum, and extraction directory.

set -e
base=https://huggingface.co/oruk/orukeet/resolve/55a984d46f68323301837194ce647c702f55facc/onnx
curl -fL "$base/manifest.json" -o orukeet-manifest.json
curl -fL "$base/sherpa-onnx-orukeet-v0.1.0-int8.tar.bz2" \
  -o sherpa-onnx-orukeet-v0.1.0-int8.tar.bz2

python3 - <<'PY'
import hashlib
import json
from pathlib import Path

manifest_path = Path("orukeet-manifest.json")
expected = "7e80f93f0e9b923c392424b0f85d28a717feee0a4d2a6aa9bfa723693868e727"
assert hashlib.sha256(manifest_path.read_bytes()).hexdigest() == expected
manifest = json.loads(manifest_path.read_text())
archive = Path(manifest["archive"])
assert archive.stat().st_size == manifest["archive_bytes"]
digest = hashlib.sha256()
with archive.open("rb") as stream:
    for block in iter(lambda: stream.read(1024 * 1024), b""):
        digest.update(block)
assert digest.hexdigest() == manifest["archive_sha256"]
print("Verified:", archive)
PY

tar xf sherpa-onnx-orukeet-v0.1.0-int8.tar.bz2

The archive contains encoder.int8.onnx, decoder.int8.onnx, joiner.int8.onnx, tokens.txt, and the weight license and attribution. Keep these files together. Downloads occupy approximately 487 MB; the extracted model occupies approximately 672 MB. Existing model files can be reused offline.

Serve

From the sherpa-onnx repository root, install the example's dependencies. This example uses the legacy WebSocket server API:

python3 -m pip install sherpa-onnx numpy 'websockets<14'

Start the server:

model=./sherpa-onnx-orukeet-v0.1.0-int8
python3 python-api-examples/non_streaming_server.py \
  --model-type nemo_transducer \
  --encoder "$model/encoder.int8.onnx" \
  --decoder "$model/decoder.int8.onnx" \
  --joiner "$model/joiner.int8.onnx" \
  --tokens "$model/tokens.txt" \
  --feat-dim 128 \
  --num-threads 2 \
  --provider cpu \
  --port 6006

Transcribe files with the existing client:

python3 python-api-examples/offline-websocket-client-decode-files-sequential.py \
  --server-addr localhost --server-port 6006 recording.wav

Audio is decoded locally. The model uses a 16 kHz frontend and the existing greedy TDT decoder. This is utterance-based recognition: the client sends each recording for a final transcript.

For microphone input, download silero_vad.onnx as described in vad-with-non-streaming-asr.py, then run:

python3 -m pip install sounddevice
model=./sherpa-onnx-orukeet-v0.1.0-int8
python3 python-api-examples/vad-with-non-streaming-asr.py \
  --silero-vad-model ./silero_vad.onnx \
  --model-type nemo_transducer \
  --encoder "$model/encoder.int8.onnx" \
  --decoder "$model/decoder.int8.onnx" \
  --joiner "$model/joiner.int8.onnx" \
  --tokens "$model/tokens.txt" \
  --feature-dim 128 --num-threads 2

Stock Parakeet TDT v3 uses the same server options with its own model directory. Other transducer examples retain the default --model-type transducer.