Export speaker verification models from NeMo to ONNX

.github/scripts/test-speaker-recognition-python.sh

@@ -1,6 +1,6 @@
 #!/usr/bin/env bash
 
-set -e
+set -ex
 
 log() {
   # This function is from espnet
@@ -21,18 +21,19 @@ model_dir=$d/wespeaker
 mkdir -p $model_dir
 pushd $model_dir
 models=(
-en_voxceleb_CAM++.onnx
-en_voxceleb_CAM++_LM.onnx
-en_voxceleb_resnet152_LM.onnx
-en_voxceleb_resnet221_LM.onnx
-en_voxceleb_resnet293_LM.onnx
-en_voxceleb_resnet34.onnx
-en_voxceleb_resnet34_LM.onnx
-zh_cnceleb_resnet34.onnx
-zh_cnceleb_resnet34_LM.onnx
+wespeaker_en_voxceleb_CAM++.onnx
+wespeaker_en_voxceleb_CAM++_LM.onnx
+wespeaker_en_voxceleb_resnet152_LM.onnx
+wespeaker_en_voxceleb_resnet221_LM.onnx
+wespeaker_en_voxceleb_resnet293_LM.onnx
+wespeaker_en_voxceleb_resnet34.onnx
+wespeaker_en_voxceleb_resnet34_LM.onnx
+wespeaker_zh_cnceleb_resnet34.onnx
+wespeaker_zh_cnceleb_resnet34_LM.onnx
 )
 for m in ${models[@]}; do
   wget -q https://github.com/k2-fsa/sherpa-onnx/releases/download/speaker-recongition-models/$m
+  wget -q https://github.com/k2-fsa/sherpa-onnx/releases/download/speaker-recongition-models/wespeaker_en_voxceleb_CAM++_LM.onnx
 done
 ls -lh
 popd
@@ -42,13 +43,13 @@ model_dir=$d/3dspeaker
 mkdir -p $model_dir
 pushd $model_dir
 models=(
-speech_campplus_sv_en_voxceleb_16k.onnx
-speech_campplus_sv_zh-cn_16k-common.onnx
-speech_eres2net_base_200k_sv_zh-cn_16k-common.onnx
-speech_eres2net_base_sv_zh-cn_3dspeaker_16k.onnx
-speech_eres2net_large_sv_zh-cn_3dspeaker_16k.onnx
-speech_eres2net_sv_en_voxceleb_16k.onnx
-speech_eres2net_sv_zh-cn_16k-common.onnx
+3dspeaker_speech_campplus_sv_en_voxceleb_16k.onnx
+3dspeaker_speech_campplus_sv_zh-cn_16k-common.onnx
+3dspeaker_speech_eres2net_base_200k_sv_zh-cn_16k-common.onnx
+3dspeaker_speech_eres2net_base_sv_zh-cn_3dspeaker_16k.onnx
+3dspeaker_speech_eres2net_large_sv_zh-cn_3dspeaker_16k.onnx
+3dspeaker_speech_eres2net_sv_en_voxceleb_16k.onnx
+3dspeaker_speech_eres2net_sv_zh-cn_16k-common.onnx
 )
 for m in ${models[@]}; do
   wget -q https://github.com/k2-fsa/sherpa-onnx/releases/download/speaker-recongition-models/$m

.github/workflows/export-nemo-speaker-verification-to-onnx.yaml

@@ -0,0 +1,45 @@
+name: export-nemo-speaker-verification-to-onnx
+
+on:
+  workflow_dispatch:
+
+concurrency:
+  group: export-nemo-speaker-verification-to-onnx-${{ github.ref }}
+  cancel-in-progress: true
+
+jobs:
+  export-nemo-speaker-verification-to-onnx:
+    if: github.repository_owner == 'k2-fsa' || github.repository_owner == 'csukuangfj'
+    name: export nemo speaker verification models to ONNX
+    runs-on: ${{ matrix.os }}
+    strategy:
+      fail-fast: false
+      matrix:
+        os: [ubuntu-latest]
+        python-version: ["3.10"]
+
+    steps:
+      - uses: actions/checkout@v4
+
+      - name: Setup Python ${{ matrix.python-version }}
+        uses: actions/setup-python@v4
+        with:
+          python-version: ${{ matrix.python-version }}
+
+      - name: Run
+        shell: bash
+        run: |
+          cd scripts/nemo/speaker-verification
+          ./run.sh
+
+          mv -v *.onnx ../../..
+
+      - name: Release
+        uses: svenstaro/upload-release-action@v2
+        with:
+          file_glob: true
+          file: ./*.onnx
+          overwrite: true
+          repo_name: k2-fsa/sherpa-onnx
+          repo_token: ${{ secrets.UPLOAD_GH_SHERPA_ONNX_TOKEN }}
+          tag: speaker-recongition-models

python-api-examples/speaker-identification.py

@@ -29,7 +29,7 @@
 https://github.com/k2-fsa/sherpa-onnx/releases/tag/speaker-recongition-models
 to download a model. An example is given below:
 
-    wget https://github.com/k2-fsa/sherpa-onnx/releases/download/speaker-recongition-models/zh_cnceleb_resnet34.onnx
+    wget https://github.com/k2-fsa/sherpa-onnx/releases/download/speaker-recongition-models/wespeaker_zh_cnceleb_resnet34.onnx
 
 Note that `zh` means Chinese, while `en` means English.
 
@@ -39,7 +39,7 @@
 
 python3 ./python-api-examples/speaker-identification.py \
   --speaker-file ./speaker.txt \
-  --model ./zh_cnceleb_resnet34.onnx
+  --model ./wespeaker_zh_cnceleb_resnet34.onnx
 """
 import argparse
 import queue

scripts/3dspeaker/run.sh

@@ -60,4 +60,6 @@ for model in ${models[@]}; do
     --model ${model}.onnx \
     --file1 ./speaker1_a_en_16k.wav \
     --file2 ./speaker2_a_en_16k.wav
+
+  mv ${model}.onnx 3dspeaker_${model}.onnx
 done

scripts/nemo/README.md

@@ -0,0 +1,7 @@
+# Introduction
+
+This directory contains scripts for exporting models
+from [NeMo](https://github.com/NVIDIA/NeMo/) to onnx
+so that you can use them in `sherpa-onnx`.
+
+- [./speaker-verification](./speaker-verification) contains models for speaker verification.

scripts/nemo/speaker-verification/README.md

@@ -0,0 +1,14 @@
+# Introduction
+
+This directory contains script for exporting speaker verification models
+from [NeMo](https://github.com/NVIDIA/NeMo/) to onnx
+so that you can use them in `sherpa-onnx`.
+
+Specifically, the following 4 models are exported to `sherpa-onnx`
+from
+[this page](https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/asr/speaker_recognition/results.html#speaker-recognition-models):
+
+  - [titanet_large](https://catalog.ngc.nvidia.com/orgs/nvidia/teams/nemo/models/titanet_large),
+  - [titanet_small](https://catalog.ngc.nvidia.com/orgs/nvidia/teams/nemo/models/titanet_small)
+  - [speakerverification_speakernet](https://ngc.nvidia.com/catalog/models/nvidia:nemo:speakerverification_speakernet)
+  - [ecapa_tdnn](https://ngc.nvidia.com/catalog/models/nvidia:nemo:ecapa_tdnn)

scripts/nemo/speaker-verification/export-onnx.py

@@ -0,0 +1,104 @@
+#!/usr/bin/env python3
+# Copyright    2024  Xiaomi Corp.        (authors: Fangjun Kuang)
+
+import argparse
+from typing import Dict
+
+import nemo.collections.asr as nemo_asr
+import onnx
+import torch
+
+
+def get_args():
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "--model",
+        type=str,
+        required=True,
+        choices=[
+            "speakerverification_speakernet",
+            "titanet_large",
+            "titanet_small",
+            "ecapa_tdnn",
+        ],
+    )
+    return parser.parse_args()
+
+
+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)
+    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():
+    args = get_args()
+    speaker_model_config = nemo_asr.models.EncDecSpeakerLabelModel.from_pretrained(
+        model_name=args.model, return_config=True
+    )
+    preprocessor_config = speaker_model_config["preprocessor"]
+
+    print(args.model)
+    print(speaker_model_config)
+    print(preprocessor_config)
+
+    assert preprocessor_config["n_fft"] == 512, preprocessor_config
+
+    assert (
+        preprocessor_config["_target_"]
+        == "nemo.collections.asr.modules.AudioToMelSpectrogramPreprocessor"
+    ), preprocessor_config
+
+    assert preprocessor_config["frame_splicing"] == 1, preprocessor_config
+
+    speaker_model = nemo_asr.models.EncDecSpeakerLabelModel.from_pretrained(
+        model_name=args.model
+    )
+    speaker_model.eval()
+    filename = f"nemo_en_{args.model}.onnx"
+    speaker_model.export(filename)
+
+    print(f"Adding metadata to {filename}")
+
+    comment = "This model is from NeMo."
+    url = {
+        "titanet_large": "https://catalog.ngc.nvidia.com/orgs/nvidia/teams/nemo/models/titanet_large",
+        "titanet_small": "https://catalog.ngc.nvidia.com/orgs/nvidia/teams/nemo/models/titanet_small",
+        "speakerverification_speakernet": "https://ngc.nvidia.com/catalog/models/nvidia:nemo:speakerverification_speakernet",
+        "ecapa_tdnn": "https://ngc.nvidia.com/catalog/models/nvidia:nemo:ecapa_tdnn",
+    }[args.model]
+
+    language = "English"
+
+    meta_data = {
+        "framework": "nemo",
+        "language": language,
+        "url": url,
+        "comment": comment,
+        "sample_rate": preprocessor_config["sample_rate"],
+        "output_dim": speaker_model_config["decoder"]["emb_sizes"],
+        "feature_normalize_type": preprocessor_config["normalize"],
+        "window_size_ms": int(float(preprocessor_config["window_size"]) * 1000),
+        "window_stride_ms": int(float(preprocessor_config["window_stride"]) * 1000),
+        "window_type": preprocessor_config["window"],  # e.g., hann
+        "feat_dim": preprocessor_config["features"],
+    }
+    print(meta_data)
+    add_meta_data(filename=filename, meta_data=meta_data)
+
+
+if __name__ == "__main__":
+    main()

scripts/nemo/speaker-verification/run.sh

@@ -0,0 +1,53 @@
+#!/usr/bin/env bash
+# Copyright    2024  Xiaomi Corp.        (authors: Fangjun Kuang)
+
+set -ex
+
+function install_nemo() {
+  curl https://bootstrap.pypa.io/get-pip.py -o get-pip.py
+  python3 get-pip.py
+
+  pip install torch==2.1.0+cpu torchaudio==2.1.0+cpu   -f https://download.pytorch.org/whl/torch_stable.html
+
+  pip install wget text-unidecode matplotlib>=3.3.2 onnx onnxruntime pybind11 Cython einops kaldi-native-fbank soundfile
+
+  sudo apt-get install -q -y sox libsndfile1 ffmpeg python3-pip
+
+  BRANCH='main'
+  python3 -m pip install git+https://github.com/NVIDIA/NeMo.git@$BRANCH#egg=nemo_toolkit[asr]
+}
+
+install_nemo
+
+model_list=(
+speakerverification_speakernet
+titanet_large
+titanet_small
+# ecapa_tdnn # causes errors, see https://github.com/NVIDIA/NeMo/issues/8168
+)
+
+for model in ${model_list[@]}; do
+  python3 ./export-onnx.py --model $model
+done
+
+ls -lh
+
+function download_test_data() {
+  wget -q https://github.com/csukuangfj/sr-data/raw/main/test/3d-speaker/speaker1_a_en_16k.wav
+  wget -q https://github.com/csukuangfj/sr-data/raw/main/test/3d-speaker/speaker1_b_en_16k.wav
+  wget -q https://github.com/csukuangfj/sr-data/raw/main/test/3d-speaker/speaker2_a_en_16k.wav
+}
+
+download_test_data
+
+for model in ${model_list[@]}; do
+  python3 ./test-onnx.py \
+    --model nemo_en_${model}.onnx \
+    --file1 ./speaker1_a_en_16k.wav \
+    --file2 ./speaker1_b_en_16k.wav
+
+  python3 ./test-onnx.py \
+    --model nemo_en_${model}.onnx \
+    --file1 ./speaker1_a_en_16k.wav \
+    --file2 ./speaker2_a_en_16k.wav
+done