EngineX-Iluvatar/enginex-mr_series-sherpa-onnx
Archived
16
0
Fork 0
Code Issues Pull Requests Actions Projects Releases Wiki Activity

Refactor exporting NeMo models (#2362)

Browse Source 
Refactors and extends model export support to include new NeMo Parakeet TDT int8 variants for English and Japanese, updating the Kotlin API, export scripts, test runners, and CI workflows.

- Added support for two new int8 model types in OfflineRecognizer.kt.
- Enhanced Python export scripts to perform dynamic quantization and metadata injection.
- Updated shell scripts and GitHub workflows to package, test, and publish int8 model artifacts.
This commit is contained in:
Fangjun Kuang
2025-07-09 16:02:12 +08:00
committed by GitHub
parent f1405779cf
commit 6122a678f5
19 changed files with 671 additions and 23 deletions
Download Patch File Download Diff File Expand all files Collapse all files

10
.github/workflows/export-nemo-fast-conformer-hybrid-transducer-ctc-non-streaming.yaml vendored
View File

@@ -61,6 +61,11 @@ jobs:
sherpa-onnx-nemo-fast-conformer-ctc-en-de-es-fr-14288
sherpa-onnx-nemo-fast-conformer-ctc-be-de-en-es-fr-hr-it-pl-ru-uk-20k
sherpa-onnx-nemo-parakeet_tdt_ctc_110m-en-36000
sherpa-onnx-nemo-fast-conformer-ctc-en-24500-int8
sherpa-onnx-nemo-fast-conformer-ctc-es-1424-int8
sherpa-onnx-nemo-fast-conformer-ctc-en-de-es-fr-14288-int8
sherpa-onnx-nemo-fast-conformer-ctc-be-de-en-es-fr-hr-it-pl-ru-uk-20k-int8
sherpa-onnx-nemo-parakeet_tdt_ctc_110m-en-36000-int8
)
for m in ${models[@]}; do
@@ -89,6 +94,11 @@ jobs:
sherpa-onnx-nemo-fast-conformer-ctc-en-de-es-fr-14288
sherpa-onnx-nemo-fast-conformer-ctc-be-de-en-es-fr-hr-it-pl-ru-uk-20k
sherpa-onnx-nemo-parakeet_tdt_ctc_110m-en-36000
sherpa-onnx-nemo-fast-conformer-ctc-en-24500-int8
sherpa-onnx-nemo-fast-conformer-ctc-es-1424-int8
sherpa-onnx-nemo-fast-conformer-ctc-en-de-es-fr-14288-int8
sherpa-onnx-nemo-fast-conformer-ctc-be-de-en-es-fr-hr-it-pl-ru-uk-20k-int8
sherpa-onnx-nemo-parakeet_tdt_ctc_110m-en-36000-int8
)
for d in ${dirs[@]}; do
tar cjvf ${d}.tar.bz2 ./$d

57
.github/workflows/export-nemo-fast-conformer-hybrid-transducer-ctc.yaml vendored
View File

@@ -54,13 +54,18 @@ jobs:
curl -SL -O https://hf-mirror.com/csukuangfj/sherpa-onnx-nemo-ctc-en-conformer-small/resolve/main/test_wavs/trans.txt
popd
cp -av test_wavs ./sherpa-onnx-nemo-streaming-fast-conformer-ctc-en-80ms
cp -av test_wavs ./sherpa-onnx-nemo-streaming-fast-conformer-ctc-en-480ms
cp -av test_wavs ./sherpa-onnx-nemo-streaming-fast-conformer-ctc-en-1040ms
tar cjvf sherpa-onnx-nemo-streaming-fast-conformer-ctc-en-80ms.tar.bz2 sherpa-onnx-nemo-streaming-fast-conformer-ctc-en-80ms
tar cjvf sherpa-onnx-nemo-streaming-fast-conformer-ctc-en-480ms.tar.bz2 sherpa-onnx-nemo-streaming-fast-conformer-ctc-en-480ms
tar cjvf sherpa-onnx-nemo-streaming-fast-conformer-ctc-en-1040ms.tar.bz2 sherpa-onnx-nemo-streaming-fast-conformer-ctc-en-1040ms
names=(
sherpa-onnx-nemo-streaming-fast-conformer-ctc-en-80ms
sherpa-onnx-nemo-streaming-fast-conformer-ctc-en-480ms
sherpa-onnx-nemo-streaming-fast-conformer-ctc-en-1040ms
sherpa-onnx-nemo-streaming-fast-conformer-ctc-en-80ms-int8
sherpa-onnx-nemo-streaming-fast-conformer-ctc-en-480ms-int8
sherpa-onnx-nemo-streaming-fast-conformer-ctc-en-1040ms-int8
)
for d in ${names[@]}; do
cp -av test_wavs $d/
tar cjvf $d.tar.bz2 $d
done
- name: Release
uses: svenstaro/upload-release-action@v2
@@ -71,3 +76,41 @@ jobs:
repo_name: k2-fsa/sherpa-onnx
repo_token: ${{ secrets.UPLOAD_GH_SHERPA_ONNX_TOKEN }}
tag: asr-models
- name: Publish to huggingface
env:
HF_TOKEN: ${{ secrets.HF_TOKEN }}
uses: nick-fields/retry@v3
with:
max_attempts: 20
timeout_seconds: 200
shell: bash
command: |
git config --global user.email "csukuangfj@gmail.com"
git config --global user.name "Fangjun Kuang"
models=(
sherpa-onnx-nemo-streaming-fast-conformer-ctc-en-80ms
sherpa-onnx-nemo-streaming-fast-conformer-ctc-en-480ms
sherpa-onnx-nemo-streaming-fast-conformer-ctc-en-1040ms
sherpa-onnx-nemo-streaming-fast-conformer-ctc-en-80ms-int8
sherpa-onnx-nemo-streaming-fast-conformer-ctc-en-480ms-int8
sherpa-onnx-nemo-streaming-fast-conformer-ctc-en-1040ms-int8
)
for m in ${models[@]}; do
rm -rf huggingface
export GIT_LFS_SKIP_SMUDGE=1
export GIT_CLONE_PROTECTION_ACTIVE=false
git clone https://csukuangfj:$HF_TOKEN@huggingface.co/csukuangfj/$m huggingface
cp -av $m/* huggingface
cd huggingface
git lfs track "*.onnx"
git lfs track "*.wav"
git status
git add .
git status
git commit -m "first commit"
git push https://csukuangfj:$HF_TOKEN@huggingface.co/csukuangfj/$m main
cd ..
done

10
.github/workflows/export-nemo-fast-conformer-hybrid-transducer-transducer-non-streaming.yaml vendored
View File

@@ -61,6 +61,11 @@ jobs:
sherpa-onnx-nemo-fast-conformer-transducer-en-de-es-fr-14288
sherpa-onnx-nemo-fast-conformer-transducer-be-de-en-es-fr-hr-it-pl-ru-uk-20k
sherpa-onnx-nemo-parakeet_tdt_transducer_110m-en-36000
sherpa-onnx-nemo-fast-conformer-transducer-en-24500-int8
sherpa-onnx-nemo-fast-conformer-transducer-es-1424-int8
sherpa-onnx-nemo-fast-conformer-transducer-en-de-es-fr-14288-int8
sherpa-onnx-nemo-fast-conformer-transducer-be-de-en-es-fr-hr-it-pl-ru-uk-20k-int8
sherpa-onnx-nemo-parakeet_tdt_transducer_110m-en-36000-int8
)
for m in ${models[@]}; do
@@ -88,6 +93,11 @@ jobs:
sherpa-onnx-nemo-fast-conformer-transducer-en-de-es-fr-14288
sherpa-onnx-nemo-fast-conformer-transducer-be-de-en-es-fr-hr-it-pl-ru-uk-20k
sherpa-onnx-nemo-parakeet_tdt_transducer_110m-en-36000
sherpa-onnx-nemo-fast-conformer-transducer-en-24500-int8
sherpa-onnx-nemo-fast-conformer-transducer-es-1424-int8
sherpa-onnx-nemo-fast-conformer-transducer-en-de-es-fr-14288-int8
sherpa-onnx-nemo-fast-conformer-transducer-be-de-en-es-fr-hr-it-pl-ru-uk-20k-int8
sherpa-onnx-nemo-parakeet_tdt_transducer_110m-en-36000-int8
)
for d in ${dirs[@]}; do
tar cjvf ${d}.tar.bz2 ./$d

57
.github/workflows/export-nemo-fast-conformer-hybrid-transducer-transducer.yaml vendored
View File

@@ -54,13 +54,18 @@ jobs:
curl -SL -O https://hf-mirror.com/csukuangfj/sherpa-onnx-nemo-ctc-en-conformer-small/resolve/main/test_wavs/trans.txt
popd
cp -av test_wavs ./sherpa-onnx-nemo-streaming-fast-conformer-transducer-en-80ms
cp -av test_wavs ./sherpa-onnx-nemo-streaming-fast-conformer-transducer-en-480ms
cp -av test_wavs ./sherpa-onnx-nemo-streaming-fast-conformer-transducer-en-1040ms
tar cjvf sherpa-onnx-nemo-streaming-fast-conformer-transducer-en-80ms.tar.bz2 sherpa-onnx-nemo-streaming-fast-conformer-transducer-en-80ms
tar cjvf sherpa-onnx-nemo-streaming-fast-conformer-transducer-en-480ms.tar.bz2 sherpa-onnx-nemo-streaming-fast-conformer-transducer-en-480ms
tar cjvf sherpa-onnx-nemo-streaming-fast-conformer-transducer-en-1040ms.tar.bz2 sherpa-onnx-nemo-streaming-fast-conformer-transducer-en-1040ms
models=(
sherpa-onnx-nemo-streaming-fast-conformer-transducer-en-80ms
sherpa-onnx-nemo-streaming-fast-conformer-transducer-en-480ms
sherpa-onnx-nemo-streaming-fast-conformer-transducer-en-1040ms
sherpa-onnx-nemo-streaming-fast-conformer-transducer-en-80ms-int8
sherpa-onnx-nemo-streaming-fast-conformer-transducer-en-480ms-int8
sherpa-onnx-nemo-streaming-fast-conformer-transducer-en-1040ms-int8
)
for m in ${models[@]}; do
cp -av test_wavs $m
tar cjvf $m.tar.bz2 $m
done
- name: Release
uses: svenstaro/upload-release-action@v2
@@ -71,3 +76,41 @@ jobs:
repo_name: k2-fsa/sherpa-onnx
repo_token: ${{ secrets.UPLOAD_GH_SHERPA_ONNX_TOKEN }}
tag: asr-models
- name: Publish to huggingface
env:
HF_TOKEN: ${{ secrets.HF_TOKEN }}
uses: nick-fields/retry@v3
with:
max_attempts: 20
timeout_seconds: 200
shell: bash
command: |
git config --global user.email "csukuangfj@gmail.com"
git config --global user.name "Fangjun Kuang"
models=(
sherpa-onnx-nemo-streaming-fast-conformer-transducer-en-80ms
sherpa-onnx-nemo-streaming-fast-conformer-transducer-en-480ms
sherpa-onnx-nemo-streaming-fast-conformer-transducer-en-1040ms
sherpa-onnx-nemo-streaming-fast-conformer-transducer-en-80ms-int8
sherpa-onnx-nemo-streaming-fast-conformer-transducer-en-480ms-int8
sherpa-onnx-nemo-streaming-fast-conformer-transducer-en-1040ms-int8
)
for m in ${models[@]}; do
rm -rf huggingface
export GIT_LFS_SKIP_SMUDGE=1
export GIT_CLONE_PROTECTION_ACTIVE=false
git clone https://csukuangfj:$HF_TOKEN@huggingface.co/csukuangfj/$m huggingface
cp -av $m/* huggingface
cd huggingface
git lfs track "*.onnx"
git lfs track "*.wav"
git status
git add .
git status
git commit -m "first commit"
git push https://csukuangfj:$HF_TOKEN@huggingface.co/csukuangfj/$m main
cd ..
done

105
.github/workflows/export-nemo-parakeet-tdt.yaml vendored Normal file
View File

@@ -0,0 +1,105 @@
name: export-nemo-parakeet-tdt
on:
push:
branches:
- refactor-export-nemo
workflow_dispatch:
concurrency:
group: export-nemo-parakeet-tdt-${{ github.ref }}
cancel-in-progress: true
jobs:
export-nemo-parakeet-tdt-0_6b-v2:
if: github.repository_owner == 'k2-fsa' || github.repository_owner == 'csukuangfj'
name: parakeet tdt
runs-on: ${{ matrix.os }}
strategy:
fail-fast: false
matrix:
os: [macos-latest]
python-version: ["3.10"]
steps:
- uses: actions/checkout@v4
- name: Setup Python ${{ matrix.python-version }}
uses: actions/setup-python@v5
with:
python-version: ${{ matrix.python-version }}
- name: Install python dependencies
shell: bash
run: |
pip install \
nemo_toolkit['asr'] \
"numpy<2" \
ipython \
kaldi-native-fbank \
librosa \
onnx==1.17.0 \
onnxmltools==1.13.0 \
onnxruntime==1.17.1 \
soundfile
- name: Run
shell: bash
run: |
cd scripts/nemo/parakeet-tdt_ctc-0.6b-ja
./run-ctc.sh
- name: Collect files
shell: bash
run: |
models=(
sherpa-onnx-nemo-parakeet-tdt_ctc-0.6b-ja-35000-int8
)
for m in ${models[@]}; do
mv -v scripts/nemo/parakeet-tdt_ctc-0.6b-ja/$m .
tar cjfv $m.tar.bz2 $m
done
- name: Publish to huggingface
env:
HF_TOKEN: ${{ secrets.HF_TOKEN }}
uses: nick-fields/retry@v3
with:
max_attempts: 20
timeout_seconds: 200
shell: bash
command: |
git config --global user.email "csukuangfj@gmail.com"
git config --global user.name "Fangjun Kuang"
models=(
sherpa-onnx-nemo-parakeet-tdt_ctc-0.6b-ja-35000-int8
)
for m in ${models[@]}; do
rm -rf huggingface
export GIT_LFS_SKIP_SMUDGE=1
export GIT_CLONE_PROTECTION_ACTIVE=false
git clone https://csukuangfj:$HF_TOKEN@huggingface.co/csukuangfj/$m huggingface
cp -av $m/* huggingface
cd huggingface
git lfs track "*.onnx"
git lfs track "*.wav"
git status
git add .
git status
git commit -m "first commit"
git push https://csukuangfj:$HF_TOKEN@huggingface.co/csukuangfj/$m main
cd ..
done
- name: Release
uses: svenstaro/upload-release-action@v2
with:
file_glob: true
file: ./*.tar.bz2
overwrite: true
repo_name: k2-fsa/sherpa-onnx
repo_token: ${{ secrets.UPLOAD_GH_SHERPA_ONNX_TOKEN }}
tag: asr-models

32
scripts/apk/generate-vad-asr-apk-script.py
View File

@@ -565,6 +565,38 @@ def get_models():
ls -lh
popd
""",
),
Model(
model_name="sherpa-onnx-nemo-parakeet_tdt_ctc_110m-en-36000-int8",
idx=33,
lang="en",
lang2="English",
short_name="parakeet_tdt_ctc_110m",
cmd="""
pushd $model_name
rm -rfv test_wavs
ls -lh
popd
""",
),
Model(
model_name="sherpa-onnx-nemo-parakeet-tdt_ctc-0.6b-ja-35000-int8",
idx=34,
lang="ja",
lang2="Japanese",
short_name="parakeet-tdt_ctc_0.6b_ja",
cmd="""
pushd $model_name
rm -rfv test_wavs
ls -lh
popd
""",
),

1
scripts/nemo/fast-conformer-hybrid-transducer-ctc/README.md
View File

@@ -23,5 +23,6 @@ This folder contains scripts for exporting models from
- https://catalog.ngc.nvidia.com/orgs/nvidia/teams/nemo/models/stt_multilingual_fastconformer_hybrid_large_pc
- https://catalog.ngc.nvidia.com/orgs/nvidia/teams/nemo/models/parakeet-tdt_ctc-110m
- https://huggingface.co/nvidia/parakeet-tdt_ctc-0.6b-ja
to `sherpa-onnx`.

7
scripts/nemo/fast-conformer-hybrid-transducer-ctc/export-onnx-ctc-non-streaming.py
View File

@@ -6,6 +6,7 @@ from typing import Dict
import nemo.collections.asr as nemo_asr
import onnx
import torch
from onnxruntime.quantization import QuantType, quantize_dynamic
def get_args():
@@ -86,6 +87,12 @@ def main():
}
add_meta_data(filename, meta_data)
quantize_dynamic(
model_input="./model.onnx",
model_output="./model.int8.onnx",
weight_type=QuantType.QUInt8,
)
print("preprocessor", asr_model.cfg.preprocessor)
print(meta_data)

6
scripts/nemo/fast-conformer-hybrid-transducer-ctc/export-onnx-ctc.py
View File

@@ -6,6 +6,7 @@ from typing import Dict
import nemo.collections.asr as nemo_asr
import onnx
import torch
from onnxruntime.quantization import QuantType, quantize_dynamic
def get_args():
@@ -114,6 +115,11 @@ def main():
"comment": "Only the CTC branch is exported",
}
add_meta_data(filename, meta_data)
quantize_dynamic(
model_input="./model.onnx",
model_output="./model.int8.onnx",
weight_type=QuantType.QUInt8,
)
print(meta_data)

8
scripts/nemo/fast-conformer-hybrid-transducer-ctc/export-onnx-transducer-non-streaming.py
View File

@@ -6,6 +6,7 @@ from typing import Dict
import nemo.collections.asr as nemo_asr
import onnx
import torch
from onnxruntime.quantization import QuantType, quantize_dynamic
def get_args():
@@ -90,6 +91,13 @@ def main():
}
add_meta_data("encoder.onnx", meta_data)
for m in ["encoder", "decoder", "joiner"]:
quantize_dynamic(
model_input=f"{m}.onnx",
model_output=f"{m}.int8.onnx",
weight_type=QuantType.QUInt8,
)
print(meta_data)

8
scripts/nemo/fast-conformer-hybrid-transducer-ctc/export-onnx-transducer.py
View File

@@ -6,6 +6,7 @@ from typing import Dict
import nemo.collections.asr as nemo_asr
import onnx
import torch
from onnxruntime.quantization import QuantType, quantize_dynamic
def get_args():
@@ -122,6 +123,13 @@ def main():
}
add_meta_data("encoder.onnx", meta_data)
for m in ["encoder", "decoder", "joiner"]:
quantize_dynamic(
model_input=f"{m}.onnx",
model_output=f"{m}.int8.onnx",
weight_type=QuantType.QUInt8,
)
print(meta_data)

77
scripts/nemo/fast-conformer-hybrid-transducer-ctc/run-ctc-non-streaming.sh
View File

@@ -19,6 +19,12 @@ log "Process $name at $url"
d=sherpa-onnx-nemo-parakeet_tdt_ctc_110m-en-36000
mkdir -p $d
mv -v model.onnx $d/
cp -v tokens.txt $d/
ls -lh $d
d=sherpa-onnx-nemo-parakeet_tdt_ctc_110m-en-36000-int8
mkdir -p $d
mv -v model.int8.onnx $d/
mv -v tokens.txt $d/
ls -lh $d
@@ -33,6 +39,12 @@ log "Process $name at $url"
d=sherpa-onnx-nemo-fast-conformer-ctc-en-24500
mkdir -p $d
mv -v model.onnx $d/
cp -v tokens.txt $d/
ls -lh $d
d=sherpa-onnx-nemo-fast-conformer-ctc-en-24500-int8
mkdir -p $d
mv -v model.int8.onnx $d/
mv -v tokens.txt $d/
ls -lh $d
@@ -45,6 +57,12 @@ doc="This collection contains the Spanish FastConformer Hybrid (CTC and Transduc
d=sherpa-onnx-nemo-fast-conformer-ctc-es-1424
mkdir -p $d
mv -v model.onnx $d/
cp -v tokens.txt $d/
ls -lh $d
d=sherpa-onnx-nemo-fast-conformer-ctc-es-1424-int8
mkdir -p $d
mv -v model.int8.onnx $d/
mv -v tokens.txt $d/
ls -lh $d
@@ -57,6 +75,12 @@ doc="This collection contains the Multilingual FastConformer Hybrid (Transducer
d=sherpa-onnx-nemo-fast-conformer-ctc-en-de-es-fr-14288
mkdir -p $d
mv -v model.onnx $d/
cp -v tokens.txt $d/
ls -lh $d
d=sherpa-onnx-nemo-fast-conformer-ctc-en-de-es-fr-14288-int8
mkdir -p $d
mv -v model.int8.onnx $d/
mv -v tokens.txt $d/
ls -lh $d
@@ -69,6 +93,12 @@ doc="This collection contains the Multilingual FastConformer Hybrid (Transducer
d=sherpa-onnx-nemo-fast-conformer-ctc-be-de-en-es-fr-hr-it-pl-ru-uk-20k
mkdir -p $d
mv -v model.onnx $d/
cp -v tokens.txt $d/
ls -lh $d
d=sherpa-onnx-nemo-fast-conformer-ctc-be-de-en-es-fr-hr-it-pl-ru-uk-20k-int8
mkdir -p $d
mv -v model.int8.onnx $d/
mv -v tokens.txt $d/
ls -lh $d
@@ -92,6 +122,16 @@ mkdir -p $d/test_wavs
cp en.wav $d/test_wavs/0.wav
cp -v $data/en-english.wav $d/test_wavs/1.wav
d=sherpa-onnx-nemo-parakeet_tdt_ctc_110m-en-36000-int8
python3 ./test-onnx-ctc-non-streaming.py \
--model $d/model.int8.onnx \
--tokens $d/tokens.txt \
--wav $data/en-english.wav
mkdir -p $d/test_wavs
cp en.wav $d/test_wavs/0.wav
cp -v $data/en-english.wav $d/test_wavs/1.wav
d=sherpa-onnx-nemo-fast-conformer-ctc-en-24500
python3 ./test-onnx-ctc-non-streaming.py \
--model $d/model.onnx \
@@ -101,6 +141,15 @@ mkdir -p $d/test_wavs
cp en.wav $d/test_wavs/0.wav
cp -v $data/en-english.wav $d/test_wavs
d=sherpa-onnx-nemo-fast-conformer-ctc-en-24500-int8
python3 ./test-onnx-ctc-non-streaming.py \
--model $d/model.int8.onnx \
--tokens $d/tokens.txt \
--wav $data/en-english.wav
mkdir -p $d/test_wavs
cp en.wav $d/test_wavs/0.wav
cp -v $data/en-english.wav $d/test_wavs
d=sherpa-onnx-nemo-fast-conformer-ctc-es-1424
python3 ./test-onnx-ctc-non-streaming.py \
--model $d/model.onnx \
@@ -109,6 +158,14 @@ python3 ./test-onnx-ctc-non-streaming.py \
mkdir -p $d/test_wavs
cp -v $data/es-spanish.wav $d/test_wavs
d=sherpa-onnx-nemo-fast-conformer-ctc-es-1424-int8
python3 ./test-onnx-ctc-non-streaming.py \
--model $d/model.int8.onnx \
--tokens $d/tokens.txt \
--wav $data/es-spanish.wav
mkdir -p $d/test_wavs
cp -v $data/es-spanish.wav $d/test_wavs
d=sherpa-onnx-nemo-fast-conformer-ctc-en-de-es-fr-14288
mkdir -p $d/test_wavs
for w in en-english.wav de-german.wav es-spanish.wav fr-french.wav; do
@@ -119,6 +176,16 @@ for w in en-english.wav de-german.wav es-spanish.wav fr-french.wav; do
cp -v $data/$w $d/test_wavs
done
d=sherpa-onnx-nemo-fast-conformer-ctc-en-de-es-fr-14288-int8
mkdir -p $d/test_wavs
for w in en-english.wav de-german.wav es-spanish.wav fr-french.wav; do
python3 ./test-onnx-ctc-non-streaming.py \
--model $d/model.int8.onnx \
--tokens $d/tokens.txt \
--wav $data/$w
cp -v $data/$w $d/test_wavs
done
d=sherpa-onnx-nemo-fast-conformer-ctc-be-de-en-es-fr-hr-it-pl-ru-uk-20k
mkdir -p $d/test_wavs
for w in en-english.wav de-german.wav es-spanish.wav fr-french.wav hr-croatian.wav it-italian.wav po-polish.wav ru-russian.wav uk-ukrainian.wav; do
@@ -128,3 +195,13 @@ for w in en-english.wav de-german.wav es-spanish.wav fr-french.wav hr-croatian.w
--wav $data/$w
cp -v $data/$w $d/test_wavs
done
d=sherpa-onnx-nemo-fast-conformer-ctc-be-de-en-es-fr-hr-it-pl-ru-uk-20k-int8
mkdir -p $d/test_wavs
for w in en-english.wav de-german.wav es-spanish.wav fr-french.wav hr-croatian.wav it-italian.wav po-polish.wav ru-russian.wav uk-ukrainian.wav; do
python3 ./test-onnx-ctc-non-streaming.py \
--model $d/model.int8.onnx \
--tokens $d/tokens.txt \
--wav $data/$w
cp -v $data/$w $d/test_wavs
done

23
scripts/nemo/fast-conformer-hybrid-transducer-ctc/run-ctc.sh
View File

@@ -17,11 +17,22 @@ ms=(
for m in ${ms[@]}; do
./export-onnx-ctc.py --model $m
d=sherpa-onnx-nemo-streaming-fast-conformer-ctc-en-${m}ms
d_int8=sherpa-onnx-nemo-streaming-fast-conformer-ctc-en-${m}ms-int8
if [ ! -f $d/model.onnx ]; then
mkdir -p $d
mkdir -p $d $d_int8
mv -v model.onnx $d/
mv -v tokens.txt $d/
cp -v tokens.txt $d/
mv -v model.int8.onnx $d_int8/
mv -v tokens.txt $d_int8/
echo "---$d---"
ls -lh $d
echo "---$d_int8---"
ls -lh $d_int8
fi
done
@@ -29,8 +40,16 @@ done
for m in ${ms[@]}; do
d=sherpa-onnx-nemo-streaming-fast-conformer-ctc-en-${m}ms
echo "---$d---"
python3 ./test-onnx-ctc.py \
--model $d/model.onnx \
--tokens $d/tokens.txt \
--wav ./0.wav
d=sherpa-onnx-nemo-streaming-fast-conformer-ctc-en-${m}ms-int8
echo "---$d---"
python3 ./test-onnx-ctc.py \
--model $d/model.int8.onnx \
--tokens $d/tokens.txt \
--wav ./0.wav
done

124
scripts/nemo/fast-conformer-hybrid-transducer-ctc/run-transducer-non-streaming.sh
View File

@@ -18,7 +18,17 @@ log "Process $name at $url"
./export-onnx-transducer-non-streaming.py --model $name --doc "$doc"
d=sherpa-onnx-nemo-parakeet_tdt_transducer_110m-en-36000
mkdir -p $d
mv -v *.onnx $d/
mv -v encoder.onnx $d/
mv -v decoder.onnx $d/
mv -v joiner.onnx $d/
cp -v tokens.txt $d/
ls -lh $d
d=sherpa-onnx-nemo-parakeet_tdt_transducer_110m-en-36000-int8
mkdir -p $d
mv -v encoder.int8.onnx $d/
mv -v decoder.int8.onnx $d/
mv -v joiner.int8.onnx $d/
mv -v tokens.txt $d/
ls -lh $d
@@ -32,7 +42,17 @@ log "Process $name at $url"
d=sherpa-onnx-nemo-fast-conformer-transducer-en-24500
mkdir -p $d
mv -v *.onnx $d/
mv -v encoder.onnx $d/
mv -v decoder.onnx $d/
mv -v joiner.onnx $d/
cp -v tokens.txt $d/
ls -lh $d
d=sherpa-onnx-nemo-fast-conformer-transducer-en-24500-int8
mkdir -p $d
mv -v encoder.int8.onnx $d/
mv -v decoder.int8.onnx $d/
mv -v joiner.int8.onnx $d/
mv -v tokens.txt $d/
ls -lh $d
@@ -44,7 +64,17 @@ doc="This collection contains the Spanish FastConformer Hybrid (CTC and Transduc
d=sherpa-onnx-nemo-fast-conformer-transducer-es-1424
mkdir -p $d
mv -v *.onnx $d/
mv -v encoder.onnx $d/
mv -v decoder.onnx $d/
mv -v joiner.onnx $d/
cp -v tokens.txt $d/
ls -lh $d
d=sherpa-onnx-nemo-fast-conformer-transducer-es-1424-int8
mkdir -p $d
mv -v encoder.int8.onnx $d/
mv -v decoder.int8.onnx $d/
mv -v joiner.int8.onnx $d/
mv -v tokens.txt $d/
ls -lh $d
@@ -56,7 +86,17 @@ doc="This collection contains the Multilingual FastConformer Hybrid (Transducer
d=sherpa-onnx-nemo-fast-conformer-transducer-en-de-es-fr-14288
mkdir -p $d
mv -v *.onnx $d/
mv -v encoder.onnx $d/
mv -v decoder.onnx $d/
mv -v joiner.onnx $d/
cp -v tokens.txt $d/
ls -lh $d
d=sherpa-onnx-nemo-fast-conformer-transducer-en-de-es-fr-14288-int8
mkdir -p $d
mv -v encoder.int8.onnx $d/
mv -v decoder.int8.onnx $d/
mv -v joiner.int8.onnx $d/
mv -v tokens.txt $d/
ls -lh $d
@@ -68,7 +108,17 @@ doc="This collection contains the Multilingual FastConformer Hybrid (Transducer
d=sherpa-onnx-nemo-fast-conformer-transducer-be-de-en-es-fr-hr-it-pl-ru-uk-20k
mkdir -p $d
mv -v *.onnx $d/
mv -v encoder.onnx $d/
mv -v decoder.onnx $d/
mv -v joiner.onnx $d/
cp -v tokens.txt $d/
ls -lh $d
d=sherpa-onnx-nemo-fast-conformer-transducer-be-de-en-es-fr-hr-it-pl-ru-uk-20k-int8
mkdir -p $d
mv -v encoder.int8.onnx $d/
mv -v decoder.int8.onnx $d/
mv -v joiner.int8.onnx $d/
mv -v tokens.txt $d/
ls -lh $d
@@ -101,6 +151,25 @@ mkdir -p $d/test_wavs
cp en.wav $d/test_wavs/0.wav
cp -v $data/en-english.wav $d/test_wavs
d=sherpa-onnx-nemo-parakeet_tdt_transducer_110m-en-36000-int8
python3 ./test-onnx-transducer-non-streaming.py \
--encoder $d/encoder.int8.onnx \
--decoder $d/decoder.int8.onnx \
--joiner $d/joiner.int8.onnx \
--tokens $d/tokens.txt \
--wav $data/en-english.wav
python3 ./test-onnx-transducer-non-streaming.py \
--encoder $d/encoder.int8.onnx \
--decoder $d/decoder.int8.onnx \
--joiner $d/joiner.int8.onnx \
--tokens $d/tokens.txt \
--wav ./en.wav
mkdir -p $d/test_wavs
cp en.wav $d/test_wavs/0.wav
cp -v $data/en-english.wav $d/test_wavs
d=sherpa-onnx-nemo-fast-conformer-transducer-en-24500
python3 ./test-onnx-transducer-non-streaming.py \
--encoder $d/encoder.onnx \
@@ -112,6 +181,17 @@ mkdir -p $d/test_wavs
cp en.wav $d/test_wavs/0.wav
cp -v $data/en-english.wav $d/test_wavs
d=sherpa-onnx-nemo-fast-conformer-transducer-en-24500-int8
python3 ./test-onnx-transducer-non-streaming.py \
--encoder $d/encoder.int8.onnx \
--decoder $d/decoder.int8.onnx \
--joiner $d/joiner.int8.onnx \
--tokens $d/tokens.txt \
--wav $data/en-english.wav
mkdir -p $d/test_wavs
cp en.wav $d/test_wavs/0.wav
cp -v $data/en-english.wav $d/test_wavs
d=sherpa-onnx-nemo-fast-conformer-transducer-es-1424
python3 ./test-onnx-transducer-non-streaming.py \
--encoder $d/encoder.onnx \
@@ -122,6 +202,16 @@ python3 ./test-onnx-transducer-non-streaming.py \
mkdir -p $d/test_wavs
cp -v $data/es-spanish.wav $d/test_wavs
d=sherpa-onnx-nemo-fast-conformer-transducer-es-1424-int8
python3 ./test-onnx-transducer-non-streaming.py \
--encoder $d/encoder.int8.onnx \
--decoder $d/decoder.int8.onnx \
--joiner $d/joiner.int8.onnx \
--tokens $d/tokens.txt \
--wav $data/es-spanish.wav
mkdir -p $d/test_wavs
cp -v $data/es-spanish.wav $d/test_wavs
d=sherpa-onnx-nemo-fast-conformer-transducer-en-de-es-fr-14288
mkdir -p $d/test_wavs
for w in en-english.wav de-german.wav es-spanish.wav fr-french.wav; do
@@ -134,6 +224,18 @@ for w in en-english.wav de-german.wav es-spanish.wav fr-french.wav; do
cp -v $data/$w $d/test_wavs
done
d=sherpa-onnx-nemo-fast-conformer-transducer-en-de-es-fr-14288-int8
mkdir -p $d/test_wavs
for w in en-english.wav de-german.wav es-spanish.wav fr-french.wav; do
python3 ./test-onnx-transducer-non-streaming.py \
--encoder $d/encoder.int8.onnx \
--decoder $d/decoder.int8.onnx \
--joiner $d/joiner.int8.onnx \
--tokens $d/tokens.txt \
--wav $data/$w
cp -v $data/$w $d/test_wavs
done
d=sherpa-onnx-nemo-fast-conformer-transducer-be-de-en-es-fr-hr-it-pl-ru-uk-20k
mkdir -p $d/test_wavs
for w in en-english.wav de-german.wav es-spanish.wav fr-french.wav hr-croatian.wav it-italian.wav po-polish.wav ru-russian.wav uk-ukrainian.wav; do
@@ -145,3 +247,15 @@ for w in en-english.wav de-german.wav es-spanish.wav fr-french.wav hr-croatian.w
--wav $data/$w
cp -v $data/$w $d/test_wavs
done
d=sherpa-onnx-nemo-fast-conformer-transducer-be-de-en-es-fr-hr-it-pl-ru-uk-20k-int8
mkdir -p $d/test_wavs
for w in en-english.wav de-german.wav es-spanish.wav fr-french.wav hr-croatian.wav it-italian.wav po-polish.wav ru-russian.wav uk-ukrainian.wav; do
python3 ./test-onnx-transducer-non-streaming.py \
--encoder $d/encoder.int8.onnx \
--decoder $d/decoder.int8.onnx \
--joiner $d/joiner.int8.onnx \
--tokens $d/tokens.txt \
--wav $data/$w
cp -v $data/$w $d/test_wavs
done

23
scripts/nemo/fast-conformer-hybrid-transducer-ctc/run-transducer.sh
View File

@@ -17,13 +17,24 @@ ms=(
for m in ${ms[@]}; do
./export-onnx-transducer.py --model $m
d=sherpa-onnx-nemo-streaming-fast-conformer-transducer-en-${m}ms
d_int8=sherpa-onnx-nemo-streaming-fast-conformer-transducer-en-${m}ms-int8
if [ ! -f $d/encoder.onnx ]; then
mkdir -p $d
mkdir -p $d $d_int8
mv -v encoder.onnx $d/
mv -v decoder.onnx $d/
mv -v joiner.onnx $d/
mv -v tokens.txt $d/
cp -v tokens.txt $d/
mv -v encoder.int8.onnx $d_int8/
mv -v decoder.int8.onnx $d_int8/
mv -v joiner.int8.onnx $d_int8/
mv -v tokens.txt $d_int8/
echo "---$d---"
ls -lh $d
echo "---$d_int8---"
ls -lh $d_int8
fi
done
@@ -37,4 +48,12 @@ for m in ${ms[@]}; do
--joiner $d/joiner.onnx \
--tokens $d/tokens.txt \
--wav ./0.wav
d=sherpa-onnx-nemo-streaming-fast-conformer-transducer-en-${m}ms-int8
python3 ./test-onnx-transducer.py \
--encoder $d/encoder.int8.onnx \
--decoder $d/decoder.int8.onnx \
--joiner $d/joiner.int8.onnx \
--tokens $d/tokens.txt \
--wav ./0.wav
done

91
scripts/nemo/parakeet-tdt_ctc-0.6b-ja/export-onnx-ctc.py Executable file
View File

@@ -0,0 +1,91 @@
#!/usr/bin/env python3
# Copyright 2025 Xiaomi Corp. (authors: Fangjun Kuang)
import os
from typing import Dict
import nemo.collections.asr as nemo_asr
import onnx
import torch
from onnxruntime.quantization import QuantType, quantize_dynamic
def add_meta_data(filename: str, meta_data: Dict[str, str]):
"""Add meta data to an ONNX model. It is changed in-place.
Args:
filename:
Filename of the ONNX model to be changed.
meta_data:
Key-value pairs.
"""
model = onnx.load(filename)
while len(model.metadata_props):
model.metadata_props.pop()
for key, value in meta_data.items():
meta = model.metadata_props.add()
meta.key = key
meta.value = str(value)
onnx.save(model, filename)
@torch.no_grad()
def main():
asr_model = nemo_asr.models.ASRModel.from_pretrained(
model_name="nvidia/parakeet-tdt_ctc-0.6b-ja"
)
print(asr_model.cfg)
print(asr_model)
with open("./tokens.txt", "w", encoding="utf-8") as f:
for i, s in enumerate(asr_model.joint.vocabulary):
f.write(f"{s} {i}\n")
f.write(f"<blk> {i+1}\n")
print("Saved to tokens.txt")
decoder_type = "ctc"
asr_model.change_decoding_strategy(decoder_type=decoder_type)
asr_model.eval()
asr_model.set_export_config({"decoder_type": "ctc"})
filename = "model.onnx"
asr_model.export(filename, onnx_opset_version=18)
normalize_type = asr_model.cfg.preprocessor.normalize
if normalize_type == "NA":
normalize_type = ""
meta_data = {
"vocab_size": asr_model.decoder.vocab_size,
"normalize_type": normalize_type,
"subsampling_factor": 8,
"model_type": "EncDecHybridRNNTCTCBPEModel",
"version": "1",
"model_author": "NeMo",
"url": "https://huggingface.co/nvidia/parakeet-tdt_ctc-0.6b-ja",
"comment": "Only the CTC branch is exported",
"doc": "See https://huggingface.co/nvidia/parakeet-tdt_ctc-0.6b-ja",
}
os.system("ls -lh *.onnx")
quantize_dynamic(
model_input="./model.onnx",
model_output="./model.int8.onnx",
weight_type=QuantType.QUInt8,
)
add_meta_data("model.int8.onnx", meta_data)
os.system("ls -lh *.onnx")
print("preprocessor", asr_model.cfg.preprocessor)
print(meta_data)
if __name__ == "__main__":
main()

34
scripts/nemo/parakeet-tdt_ctc-0.6b-ja/run-ctc.sh Executable file
View File

@@ -0,0 +1,34 @@
#!/usr/bin/env bash
set -ex
python3 ./export-onnx-ctc.py
ls -lh *.onnx
mkdir -p test_wavs
pushd test_wavs
curl -SL -O https://huggingface.co/csukuangfj/reazonspeech-k2-v2-ja-en/resolve/main/test_wavs/transcripts.txt
curl -SL -O https://hf-mirror.com/csukuangfj/reazonspeech-k2-v2-ja-en/resolve/main/test_wavs/test_ja_1.wav
curl -SL -O https://hf-mirror.com/csukuangfj/reazonspeech-k2-v2-ja-en/resolve/main/test_wavs/test_ja_2.wav
popd
d=sherpa-onnx-nemo-parakeet-tdt_ctc-0.6b-ja-35000-int8
mkdir -p $d
mv -v model.int8.onnx $d/
cp -v tokens.txt $d/
cp -av test_wavs $d
ls -lh $d
d=sherpa-onnx-nemo-parakeet-tdt_ctc-0.6b-ja-35000-int8
python3 ./test-onnx-ctc-non-streaming.py \
--model $d/model.int8.onnx \
--tokens $d/tokens.txt \
--wav $d/test_wavs/test_ja_1.wav
python3 ./test-onnx-ctc-non-streaming.py \
--model $d/model.int8.onnx \
--tokens $d/tokens.txt \
--wav $d/test_wavs/test_ja_2.wav

1
scripts/nemo/parakeet-tdt_ctc-0.6b-ja/test-onnx-ctc-non-streaming.py Symbolic link
View File

@@ -0,0 +1 @@
../fast-conformer-hybrid-transducer-ctc/test-onnx-ctc-non-streaming.py

20
sherpa-onnx/kotlin-api/OfflineRecognizer.kt
View File

@@ -601,6 +601,26 @@ fun getOfflineModelConfig(type: Int): OfflineModelConfig? {
tokens = "$modelDir/tokens.txt",
)
}
33 -> {
val modelDir = "sherpa-onnx-nemo-parakeet_tdt_ctc_110m-en-36000-int8"
return OfflineModelConfig(
nemo = OfflineNemoEncDecCtcModelConfig(
model = "$modelDir/model.int8.onnx",
),
tokens = "$modelDir/tokens.txt",
)
}
34 -> {
val modelDir = "sherpa-onnx-nemo-parakeet-tdt_ctc-0.6b-ja-35000-int8"
return OfflineModelConfig(
nemo = OfflineNemoEncDecCtcModelConfig(
model = "$modelDir/model.int8.onnx",
),
tokens = "$modelDir/tokens.txt",
)
}
}
return null
}