Add GigaAM NeMo transducer model for Russian ASR

.github/workflows/export-nemo-giga-am-to-onnx.yaml

@@ -38,7 +38,7 @@ jobs:
           mkdir $d/test_wavs
           rm scripts/nemo/GigaAM/model.onnx
           mv -v scripts/nemo/GigaAM/*.int8.onnx $d/
-          mv -v scripts/nemo/GigaAM/*.md $d/
+          cp -v scripts/nemo/GigaAM/*.md $d/
           mv -v scripts/nemo/GigaAM/*.pdf $d/
           mv -v scripts/nemo/GigaAM/tokens.txt $d/
           mv -v scripts/nemo/GigaAM/*.wav $d/test_wavs/
@@ -51,6 +51,34 @@ jobs:
 
           tar cjvf ${d}.tar.bz2 $d
 
+      - name: Run Transducer
+        shell: bash
+        run: |
+          pushd scripts/nemo/GigaAM
+          ./run-rnnt.sh
+          popd
+
+          d=sherpa-onnx-nemo-transducer-giga-am-russian-2024-10-24
+          mkdir $d
+          mkdir $d/test_wavs
+
+          mv -v scripts/nemo/GigaAM/encoder.int8.onnx $d/
+          mv -v scripts/nemo/GigaAM/decoder.onnx $d/
+          mv -v scripts/nemo/GigaAM/joiner.onnx $d/
+
+          cp -v scripts/nemo/GigaAM/*.md $d/
+          mv -v scripts/nemo/GigaAM/*.pdf $d/
+          mv -v scripts/nemo/GigaAM/tokens.txt $d/
+          mv -v scripts/nemo/GigaAM/*.wav $d/test_wavs/
+          mv -v scripts/nemo/GigaAM/run-rnnt.sh $d/
+          mv -v scripts/nemo/GigaAM/*-rnnt.py $d/
+
+          ls -lh scripts/nemo/GigaAM/
+
+          ls -lh $d
+
+          tar cjvf ${d}.tar.bz2 $d
+
       - name: Release
         uses: svenstaro/upload-release-action@v2
         with:
@@ -61,7 +89,7 @@ jobs:
           repo_token: ${{ secrets.UPLOAD_GH_SHERPA_ONNX_TOKEN }}
           tag: asr-models
 
-      - name: Publish to huggingface (CTC)
+      - name: Publish to huggingface (Transducer)
         env:
           HF_TOKEN: ${{ secrets.HF_TOKEN }}
         uses: nick-fields/retry@v3
@@ -73,7 +101,7 @@ jobs:
             git config --global user.email "[email protected]"
             git config --global user.name "Fangjun Kuang"
 
-            d=sherpa-onnx-nemo-ctc-giga-am-russian-2024-10-24
+            d=sherpa-onnx-nemo-transducer-giga-am-russian-2024-10-24/
             export GIT_LFS_SKIP_SMUDGE=1
             export GIT_CLONE_PROTECTION_ACTIVE=false
             git clone https://csukuangfj:[email protected]/csukuangfj/$d huggingface

scripts/apk/generate-vad-asr-apk-script.py

@@ -351,6 +351,24 @@ def get_models():
 
             ls -lh
 
+            popd
+            """,
+        ),
+        Model(
+            model_name="sherpa-onnx-nemo-transducer-giga-am-russian-2024-10-24",
+            idx=20,
+            lang="ru",
+            short_name="nemo_transducer_giga_am",
+            cmd="""
+            pushd $model_name
+
+            rm -rfv test_wavs
+
+            rm -fv *.sh
+            rm -fv *.py
+
+            ls -lh
+
             popd
             """,
         ),

scripts/nemo/GigaAM/export-onnx-ctc.py

@@ -75,6 +75,7 @@ def add_meta_data(filename: str, meta_data: Dict[str, str]):
     onnx.save(model, filename)
 
 
+@torch.no_grad()
 def main():
     model = EncDecCTCModel.from_config_file("./ctc_model_config.yaml")
     ckpt = torch.load("./ctc_model_weights.ckpt", map_location="cpu")

scripts/nemo/GigaAM/export-onnx-rnnt.py

@@ -0,0 +1,119 @@
+#!/usr/bin/env python3
+# Copyright      2024  Xiaomi Corp.        (authors: Fangjun Kuang)
+
+from typing import Dict
+
+import onnx
+import torch
+import torchaudio
+from nemo.collections.asr.models import EncDecRNNTBPEModel
+from nemo.collections.asr.modules.audio_preprocessing import (
+    AudioToMelSpectrogramPreprocessor as NeMoAudioToMelSpectrogramPreprocessor,
+)
+from nemo.collections.asr.parts.preprocessing.features import (
+    FilterbankFeaturesTA as NeMoFilterbankFeaturesTA,
+)
+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)
+
+
+class FilterbankFeaturesTA(NeMoFilterbankFeaturesTA):
+    def __init__(self, mel_scale: str = "htk", wkwargs=None, **kwargs):
+        if "window_size" in kwargs:
+            del kwargs["window_size"]
+        if "window_stride" in kwargs:
+            del kwargs["window_stride"]
+
+        super().__init__(**kwargs)
+
+        self._mel_spec_extractor: torchaudio.transforms.MelSpectrogram = (
+            torchaudio.transforms.MelSpectrogram(
+                sample_rate=self._sample_rate,
+                win_length=self.win_length,
+                hop_length=self.hop_length,
+                n_mels=kwargs["nfilt"],
+                window_fn=self.torch_windows[kwargs["window"]],
+                mel_scale=mel_scale,
+                norm=kwargs["mel_norm"],
+                n_fft=kwargs["n_fft"],
+                f_max=kwargs.get("highfreq", None),
+                f_min=kwargs.get("lowfreq", 0),
+                wkwargs=wkwargs,
+            )
+        )
+
+
+class AudioToMelSpectrogramPreprocessor(NeMoAudioToMelSpectrogramPreprocessor):
+    def __init__(self, mel_scale: str = "htk", **kwargs):
+        super().__init__(**kwargs)
+        kwargs["nfilt"] = kwargs["features"]
+        del kwargs["features"]
+        self.featurizer = (
+            FilterbankFeaturesTA(  # Deprecated arguments; kept for config compatibility
+                mel_scale=mel_scale,
+                **kwargs,
+            )
+        )
+
+
+@torch.no_grad()
+def main():
+    model = EncDecRNNTBPEModel.from_config_file("./rnnt_model_config.yaml")
+    ckpt = torch.load("./rnnt_model_weights.ckpt", map_location="cpu")
+    model.load_state_dict(ckpt, strict=False)
+    model.eval()
+
+    with open("./tokens.txt", "w", encoding="utf-8") as f:
+        for i, s in enumerate(model.joint.vocabulary):
+            f.write(f"{s} {i}\n")
+        f.write(f"<blk> {i+1}\n")
+        print("Saved to tokens.txt")
+
+    model.encoder.export("encoder.onnx")
+    model.decoder.export("decoder.onnx")
+    model.joint.export("joiner.onnx")
+
+    meta_data = {
+        "vocab_size": model.decoder.vocab_size,  # not including the blank
+        "pred_rnn_layers": model.decoder.pred_rnn_layers,
+        "pred_hidden": model.decoder.pred_hidden,
+        "normalize_type": "",
+        "subsampling_factor": 4,
+        "model_type": "EncDecRNNTBPEModel",
+        "version": "1",
+        "model_author": "https://github.com/salute-developers/GigaAM",
+        "license": "https://github.com/salute-developers/GigaAM/blob/main/GigaAM%20License_NC.pdf",
+        "language": "Russian",
+        "is_giga_am": 1,
+    }
+    add_meta_data("encoder.onnx", meta_data)
+
+    quantize_dynamic(
+        model_input="encoder.onnx",
+        model_output="encoder.int8.onnx",
+        weight_type=QuantType.QUInt8,
+    )
+
+
+if __name__ == "__main__":
+    main()

scripts/nemo/GigaAM/run-ctc.sh

@@ -21,11 +21,15 @@ function install_nemo() {
 }
 
 function download_files() {
-  curl -SL -O https://n-ws-q0bez.s3pd12.sbercloud.ru/b-ws-q0bez-jpv/GigaAM/ctc_model_weights.ckpt
-  curl -SL -O https://n-ws-q0bez.s3pd12.sbercloud.ru/b-ws-q0bez-jpv/GigaAM/ctc_model_config.yaml
-  curl -SL -O https://n-ws-q0bez.s3pd12.sbercloud.ru/b-ws-q0bez-jpv/GigaAM/example.wav
-  curl -SL -O https://n-ws-q0bez.s3pd12.sbercloud.ru/b-ws-q0bez-jpv/GigaAM/long_example.wav
-  curl -SL -O https://huggingface.co/csukuangfj/tmp-files/resolve/main/GigaAM%20License_NC.pdf
+  # curl -SL -O https://n-ws-q0bez.s3pd12.sbercloud.ru/b-ws-q0bez-jpv/GigaAM/ctc_model_weights.ckpt
+  # curl -SL -O https://n-ws-q0bez.s3pd12.sbercloud.ru/b-ws-q0bez-jpv/GigaAM/ctc_model_config.yaml
+  # curl -SL -O https://n-ws-q0bez.s3pd12.sbercloud.ru/b-ws-q0bez-jpv/GigaAM/example.wav
+  # curl -SL -O https://n-ws-q0bez.s3pd12.sbercloud.ru/b-ws-q0bez-jpv/GigaAM/long_example.wav
+  curl -SL -O https://huggingface.co/csukuangfj/tmp-files/resolve/main/GigaAM/ctc/ctc_model_weights.ckpt
+  curl -SL -O https://huggingface.co/csukuangfj/tmp-files/resolve/main/GigaAM/ctc/ctc_model_config.yaml
+  curl -SL -O https://huggingface.co/csukuangfj/tmp-files/resolve/main/GigaAM/example.wav
+  curl -SL -O https://huggingface.co/csukuangfj/tmp-files/resolve/main/GigaAM/long_example.wav
+  curl -SL -O https://huggingface.co/csukuangfj/tmp-files/resolve/main/GigaAM/GigaAM%20License_NC.pdf
 }
 
 install_nemo

scripts/nemo/GigaAM/run-rnnt.sh

@@ -0,0 +1,50 @@
+#!/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==2.4.0 torchaudio==2.4.0 -f https://download.pytorch.org/whl/torch_stable.html
+
+  pip install -qq wget text-unidecode matplotlib>=3.3.2 onnx onnxruntime pybind11 Cython einops kaldi-native-fbank soundfile librosa
+  pip install -qq ipython
+
+  # sudo apt-get install -q -y sox libsndfile1 ffmpeg python3-pip ipython
+
+  BRANCH='main'
+  python3 -m pip install git+https://github.com/NVIDIA/NeMo.git@$BRANCH#egg=nemo_toolkit[asr]
+
+  pip install numpy==1.26.4
+}
+
+function download_files() {
+  # curl -SL -O https://n-ws-q0bez.s3pd12.sbercloud.ru/b-ws-q0bez-jpv/GigaAM/rnnt_model_weights.ckpt
+  # curl -SL -O https://n-ws-q0bez.s3pd12.sbercloud.ru/b-ws-q0bez-jpv/GigaAM/rnnt_model_config.yaml
+  # curl -SL -O https://n-ws-q0bez.s3pd12.sbercloud.ru/b-ws-q0bez-jpv/GigaAM/example.wav
+  # curl -SL -O https://n-ws-q0bez.s3pd12.sbercloud.ru/b-ws-q0bez-jpv/GigaAM/long_example.wav
+  # curl -SL -O https://n-ws-q0bez.s3pd12.sbercloud.ru/b-ws-q0bez-jpv/GigaAM/tokenizer_all_sets.tar
+
+  curl -SL -O https://huggingface.co/csukuangfj/tmp-files/resolve/main/GigaAM/rnnt/rnnt_model_weights.ckpt
+  curl -SL -O https://huggingface.co/csukuangfj/tmp-files/resolve/main/GigaAM/rnnt/rnnt_model_config.yaml
+  curl -SL -O https://huggingface.co/csukuangfj/tmp-files/resolve/main/GigaAM/example.wav
+  curl -SL -O https://huggingface.co/csukuangfj/tmp-files/resolve/main/GigaAM/long_example.wav
+  curl -SL -O https://huggingface.co/csukuangfj/tmp-files/resolve/main/GigaAM/GigaAM%20License_NC.pdf
+  curl -SL -O https://huggingface.co/csukuangfj/tmp-files/resolve/main/GigaAM/rnnt/tokenizer_all_sets.tar
+  tar -xf tokenizer_all_sets.tar && rm tokenizer_all_sets.tar
+  ls -lh
+  echo "---"
+  ls -lh tokenizer_all_sets
+  echo "---"
+}
+
+install_nemo
+download_files
+
+python3 ./export-onnx-rnnt.py
+ls -lh
+python3 ./test-onnx-rnnt.py
+rm -v encoder.onnx
+ls -lh