Export speaker verification models from NeMo to ONNX (#526)

 #!/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

+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

@@ -29,7 +29,7 @@ Please visit
 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 @@ Assume the filename of the text file is speaker.txt.
 
 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

@@ -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

+# 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.

+# 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)

+#!/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()

+#!/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

+#!/usr/bin/env python3
+# Copyright      2023-2024  Xiaomi Corp.        (authors: Fangjun Kuang)
+
+"""
+This script computes speaker similarity score in the range [0-1]
+of two wave files using a speaker embedding model.
+"""
+import argparse
+import wave
+from pathlib import Path
+
+import kaldi_native_fbank as knf
+import numpy as np
+import onnxruntime as ort
+from numpy.linalg import norm
+
+
+def get_args():
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "--model",
+        type=str,
+        required=True,
+        help="Path to the input onnx model. Example value: model.onnx",
+    )
+
+    parser.add_argument(
+        "--file1",
+        type=str,
+        required=True,
+        help="Input wave 1",
+    )
+
+    parser.add_argument(
+        "--file2",
+        type=str,
+        required=True,
+        help="Input wave 2",
+    )
+
+    return parser.parse_args()
+
+
+def read_wavefile(filename, expected_sample_rate: int = 16000) -> np.ndarray:
+    """
+    Args:
+      filename:
+        Path to a wave file, which must be of 16-bit and 16kHz.
+     expected_sample_rate:
+       Expected sample rate of the wave file.
+    Returns:
+      Return a 1-D float32 array containing audio samples. Each sample is in
+      the range [-1, 1].
+    """
+    filename = str(filename)
+    with wave.open(filename) as f:
+        wave_file_sample_rate = f.getframerate()
+        assert wave_file_sample_rate == expected_sample_rate, (
+            wave_file_sample_rate,
+            expected_sample_rate,
+        )
+
+        num_channels = f.getnchannels()
+        assert f.getsampwidth() == 2, f.getsampwidth()  # it is in bytes
+        num_samples = f.getnframes()
+        samples = f.readframes(num_samples)
+        samples_int16 = np.frombuffer(samples, dtype=np.int16)
+        samples_int16 = samples_int16.reshape(-1, num_channels)[:, 0]
+        samples_float32 = samples_int16.astype(np.float32)
+
+        samples_float32 = samples_float32 / 32768
+
+        return samples_float32
+
+
+def compute_features(samples: np.ndarray, model: "OnnxModel") -> np.ndarray:
+    fbank_opts = knf.FbankOptions()
+    fbank_opts.frame_opts.samp_freq = model.sample_rate
+    fbank_opts.frame_opts.frame_length_ms = model.window_size_ms
+    fbank_opts.frame_opts.frame_shift_ms = model.window_stride_ms
+    fbank_opts.frame_opts.dither = 0
+    fbank_opts.frame_opts.remove_dc_offset = False
+    fbank_opts.frame_opts.window_type = model.window_type
+
+    fbank_opts.mel_opts.num_bins = model.feat_dim
+    fbank_opts.mel_opts.low_freq = 0
+    fbank_opts.mel_opts.is_librosa = True
+
+    fbank = knf.OnlineFbank(fbank_opts)
+    fbank.accept_waveform(model.sample_rate, samples)
+    fbank.input_finished()
+
+    features = []
+    for i in range(fbank.num_frames_ready):
+        f = fbank.get_frame(i)
+        features.append(f)
+    features = np.stack(features, axis=0)
+    # at this point, the shape of features is (T, C)
+
+    if model.feature_normalize_type != "":
+        assert model.feature_normalize_type == "per_feature"
+        mean = np.mean(features, axis=0, keepdims=True)
+        std = np.std(features, axis=0, keepdims=True)
+        features = (features - mean) / std
+
+    feature_len = features.shape[0]
+    pad = 16 - feature_len % 16
+
+    if pad > 0:
+        padding = np.zeros((pad, features.shape[1]), dtype=np.float32)
+        features = np.concatenate([features, padding])
+
+    features = np.expand_dims(features, axis=0)
+
+    return features, feature_len
+
+
+class OnnxModel:
+    def __init__(
+        self,
+        filename: str,
+    ):
+        session_opts = ort.SessionOptions()
+        session_opts.inter_op_num_threads = 1
+        session_opts.intra_op_num_threads = 1
+
+        self.session_opts = session_opts
+
+        self.model = ort.InferenceSession(
+            filename,
+            sess_options=self.session_opts,
+        )
+
+        meta = self.model.get_modelmeta().custom_metadata_map
+        self.framework = meta["framework"]
+        self.sample_rate = int(meta["sample_rate"])
+        self.output_dim = int(meta["output_dim"])
+        self.feature_normalize_type = meta["feature_normalize_type"]
+        self.window_size_ms = int(meta["window_size_ms"])
+        self.window_stride_ms = int(meta["window_stride_ms"])
+        self.window_type = meta["window_type"]
+        self.feat_dim = int(meta["feat_dim"])
+        print(meta)
+
+        assert self.framework == "nemo", self.framework
+
+    def __call__(self, x: np.ndarray, x_lens: int) -> np.ndarray:
+        """
+        Args:
+          x:
+            A 2-D float32 tensor of shape (T, C).
+          y:
+            A 1-D float32 tensor containing model output.
+        """
+        x = x.transpose(0, 2, 1)  # (B, T, C) -> (B, C, T)
+        x_lens = np.asarray([x_lens], dtype=np.int64)
+
+        return self.model.run(
+            [
+                self.model.get_outputs()[1].name,
+            ],
+            {
+                self.model.get_inputs()[0].name: x,
+                self.model.get_inputs()[1].name: x_lens,
+            },
+        )[0][0]
+
+
+def main():
+    args = get_args()
+    print(args)
+    filename = Path(args.model)
+    file1 = Path(args.file1)
+    file2 = Path(args.file2)
+    assert filename.is_file(), filename
+    assert file1.is_file(), file1
+    assert file2.is_file(), file2
+
+    model = OnnxModel(filename)
+    wave1 = read_wavefile(file1, model.sample_rate)
+    wave2 = read_wavefile(file2, model.sample_rate)
+
+    features1, features1_len = compute_features(wave1, model)
+    features2, features2_len = compute_features(wave2, model)
+
+    output1 = model(features1, features1_len)
+    output2 = model(features2, features2_len)
+
+    similarity = np.dot(output1, output2) / (norm(output1) * norm(output2))
+    print(f"similarity in the range [0-1]: {similarity}")
+
+
+if __name__ == "__main__":
+    main()

@@ -24,7 +24,7 @@ ls -lh
   --file1 ./wespeaker-models/test_wavs/00001_spk1.wav \
   --file2 ./wespeaker-models/test_wavs/00010_spk2.wav
 
-mv voxceleb_resnet34.onnx en_voxceleb_resnet34.onnx
+mv voxceleb_resnet34.onnx wespeaker_en_voxceleb_resnet34.onnx
 
 ./add_meta_data.py \
   --model ./voxceleb_resnet34_LM.onnx \
@@ -38,7 +38,7 @@ mv voxceleb_resnet34.onnx en_voxceleb_resnet34.onnx
   --file1 ./wespeaker-models/test_wavs/00001_spk1.wav \
   --file2 ./wespeaker-models/test_wavs/00010_spk2.wav
 
-mv voxceleb_resnet34_LM.onnx en_voxceleb_resnet34_LM.onnx
+mv voxceleb_resnet34_LM.onnx wespeaker_en_voxceleb_resnet34_LM.onnx
 
 ./add_meta_data.py \
   --model ./voxceleb_resnet152_LM.onnx \
@@ -53,7 +53,7 @@ mv voxceleb_resnet34_LM.onnx en_voxceleb_resnet34_LM.onnx
   --file1 ./wespeaker-models/test_wavs/00001_spk1.wav \
   --file2 ./wespeaker-models/test_wavs/00010_spk2.wav
 
-mv voxceleb_resnet152_LM.onnx en_voxceleb_resnet152_LM.onnx
+mv voxceleb_resnet152_LM.onnx wespeaker_en_voxceleb_resnet152_LM.onnx
 
 ./add_meta_data.py \
   --model ./voxceleb_resnet221_LM.onnx \
@@ -68,7 +68,7 @@ mv voxceleb_resnet152_LM.onnx en_voxceleb_resnet152_LM.onnx
   --file1 ./wespeaker-models/test_wavs/00001_spk1.wav \
   --file2 ./wespeaker-models/test_wavs/00010_spk2.wav
 
-mv voxceleb_resnet221_LM.onnx en_voxceleb_resnet221_LM.onnx
+mv voxceleb_resnet221_LM.onnx wespeaker_en_voxceleb_resnet221_LM.onnx
 
 ./add_meta_data.py \
   --model ./voxceleb_resnet293_LM.onnx \
@@ -83,7 +83,7 @@ mv voxceleb_resnet221_LM.onnx en_voxceleb_resnet221_LM.onnx
   --file1 ./wespeaker-models/test_wavs/00001_spk1.wav \
   --file2 ./wespeaker-models/test_wavs/00010_spk2.wav
 
-mv voxceleb_resnet293_LM.onnx en_voxceleb_resnet293_LM.onnx
+mv voxceleb_resnet293_LM.onnx wespeaker_en_voxceleb_resnet293_LM.onnx
 
 ./add_meta_data.py \
   --model ./voxceleb_CAM++.onnx \
@@ -98,7 +98,7 @@ mv voxceleb_resnet293_LM.onnx en_voxceleb_resnet293_LM.onnx
   --file1 ./wespeaker-models/test_wavs/00001_spk1.wav \
   --file2 ./wespeaker-models/test_wavs/00010_spk2.wav
 
-mv voxceleb_CAM++.onnx en_voxceleb_CAM++.onnx
+mv voxceleb_CAM++.onnx wespeaker_en_voxceleb_CAM++.onnx
 
 ./add_meta_data.py \
   --model ./voxceleb_CAM++_LM.onnx \
@@ -113,20 +113,20 @@ mv voxceleb_CAM++.onnx en_voxceleb_CAM++.onnx
   --file1 ./wespeaker-models/test_wavs/00001_spk1.wav \
   --file2 ./wespeaker-models/test_wavs/00010_spk2.wav
 
-mv voxceleb_CAM++_LM.onnx en_voxceleb_CAM++_LM.onnx
+mv voxceleb_CAM++_LM.onnx wespeaker_en_voxceleb_CAM++_LM.onnx
 
 ./add_meta_data.py \
   --model ./cnceleb_resnet34.onnx \
   --language Chinese \
   --url https://wespeaker-1256283475.cos.ap-shanghai.myqcloud.com/models/cnceleb/cnceleb_resnet34.onnx
 
-mv cnceleb_resnet34.onnx zh_cnceleb_resnet34.onnx
+mv cnceleb_resnet34.onnx wespeaker_zh_cnceleb_resnet34.onnx
 
 ./add_meta_data.py \
   --model ./cnceleb_resnet34_LM.onnx \
   --language Chinese \
   --url https://wespeaker-1256283475.cos.ap-shanghai.myqcloud.com/models/cnceleb/cnceleb_resnet34_LM.onnx
 
-mv cnceleb_resnet34_LM.onnx zh_cnceleb_resnet34_LM.onnx
+mv cnceleb_resnet34_LM.onnx wespeaker_zh_cnceleb_resnet34_LM.onnx
 
 ls -lh