Remove the 30-second constraint from whisper. (#471)

.github/scripts/test-offline-whisper.sh

@@ -16,8 +16,12 @@ which $EXE
 names=(
 tiny.en
 base.en
-# small.en
-# medium.en
+small.en
+medium.en
+tiny
+base
+small
+medium
 )
 
 for name in ${names[@]}; do
@@ -33,8 +37,8 @@ for name in ${names[@]}; do
   GIT_LFS_SKIP_SMUDGE=1 git clone $repo_url
   pushd $repo
   git lfs pull --include "*.onnx"
-  git lfs pull --include "*.ort"
-  ls -lh *.{onnx,ort}
+  # git lfs pull --include "*.ort"
+  ls -lh *.onnx
   popd
 
   log "test fp32 onnx"
@@ -43,6 +47,7 @@ for name in ${names[@]}; do
     --tokens=$repo/${name}-tokens.txt \
     --whisper-encoder=$repo/${name}-encoder.onnx \
     --whisper-decoder=$repo/${name}-decoder.onnx \
+    --whisper-tail-paddings=500 \
     --num-threads=2 \
     $repo/test_wavs/0.wav \
     $repo/test_wavs/1.wav \
@@ -54,28 +59,7 @@ for name in ${names[@]}; do
     --tokens=$repo/${name}-tokens.txt \
     --whisper-encoder=$repo/${name}-encoder.int8.onnx \
     --whisper-decoder=$repo/${name}-decoder.int8.onnx \
-    --num-threads=2 \
-    $repo/test_wavs/0.wav \
-    $repo/test_wavs/1.wav \
-    $repo/test_wavs/8k.wav
-
-  log "test fp32 ort"
-
-  time $EXE \
-    --tokens=$repo/${name}-tokens.txt \
-    --whisper-encoder=$repo/${name}-encoder.ort \
-    --whisper-decoder=$repo/${name}-decoder.ort \
-    --num-threads=2 \
-    $repo/test_wavs/0.wav \
-    $repo/test_wavs/1.wav \
-    $repo/test_wavs/8k.wav
-
-  log "test int8 ort"
-
-  time $EXE \
-    --tokens=$repo/${name}-tokens.txt \
-    --whisper-encoder=$repo/${name}-encoder.int8.ort \
-    --whisper-decoder=$repo/${name}-decoder.int8.ort \
+    --whisper-tail-paddings=500 \
     --num-threads=2 \
     $repo/test_wavs/0.wav \
     $repo/test_wavs/1.wav \

.github/workflows/export-whisper-to-onnx.yaml

@@ -15,7 +15,7 @@ jobs:
     strategy:
       fail-fast: false
       matrix:
-        os: [macos-latest]
+        os: [ubuntu-latest]
         model: ["distil-medium.en", "tiny.en", "base.en", "small.en", "medium.en", "tiny", "base", "small", "medium", "large", "large-v1", "large-v2"]
         python-version: ["3.8"]
 
@@ -44,49 +44,69 @@ jobs:
             ls -lh
           fi
           python3 ./export-onnx.py --model ${{ matrix.model }}
-          python3 -m onnxruntime.tools.convert_onnx_models_to_ort --optimization_style=Fixed ./
+          # python3 -m onnxruntime.tools.convert_onnx_models_to_ort --optimization_style=Fixed ./
 
           ls -lh
 
           if [[ $model != distil-medium.en ]]; then
             ls -lh ~/.cache/whisper
           fi
 
+          src=sherpa-onnx-whisper-${{ matrix.model }}
+
+          mkdir $src
+          cp *.onnx $src/
+          cp *tokens.txt $src
+
+          cd $src
+          mkdir -p test_wavs
+          cd test_wavs
+          wget -q https://huggingface.co/csukuangfj/sherpa-onnx-whisper-medium.en/resolve/main/test_wavs/0.wav
+          wget -q https://huggingface.co/csukuangfj/sherpa-onnx-whisper-medium.en/resolve/main/test_wavs/1.wav
+          wget -q https://huggingface.co/csukuangfj/sherpa-onnx-whisper-medium.en/resolve/main/test_wavs/8k.wav
+          wget -q https://huggingface.co/csukuangfj/sherpa-onnx-whisper-medium.en/resolve/main/test_wavs/trans.txt
+          cd ../..
+          mv $src ../..
+
+          cd ../..
+          echo "--------------------"
+          ls -lh
+          ls -lh $src
+          echo "--------------------"
+
+          tar cjvf ./$src.tar.bz2 $src
+
+      - 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
+
       - name: Publish ${{ matrix.model }} to huggingface
         shell: bash
         env:
           HF_TOKEN: ${{ secrets.HF_TOKEN }}
         run: |
-          model=${{ matrix.model }}
-
-          cd scripts/whisper
+          src=sherpa-onnx-whisper-${{ matrix.model }}
 
           git config --global user.email "[email protected]"
           git config --global user.name "Fangjun Kuang"
 
           GIT_LFS_SKIP_SMUDGE=1 git clone https://huggingface.co/csukuangfj/sherpa-onnx-whisper-${{ matrix.model }} huggingface
+          rm -rf huggingface/*
 
-          cp *.onnx ./huggingface
-          cp *.ort ./huggingface
-          cp *tokens.txt ./huggingface
+          cp -av $src/* ./huggingface/
 
           cd huggingface
 
-          if [[ $model == distil-medium.en ]]; then
-            mkdir test_wavs
-            cd test_wavs
-            wget -q https://huggingface.co/csukuangfj/sherpa-onnx-whisper-medium.en/resolve/main/test_wavs/0.wav
-            wget -q https://huggingface.co/csukuangfj/sherpa-onnx-whisper-medium.en/resolve/main/test_wavs/1.wav
-            wget -q https://huggingface.co/csukuangfj/sherpa-onnx-whisper-medium.en/resolve/main/test_wavs/8k.wav
-            wget -q https://huggingface.co/csukuangfj/sherpa-onnx-whisper-medium.en/resolve/main/test_wavs/trans.txt
-            git add .
-            cd ..
-          fi
-
           git status
           ls -lh
           git lfs track "*.onnx"
-          git lfs track "*.ort"
+          # git lfs track "*.ort"
           git add .
           git commit -m "upload ${{ matrix.model }}"
           git push https://csukuangfj:[email protected]/csukuangfj/sherpa-onnx-whisper-${{ matrix.model }} main

.github/workflows/linux.yaml

@@ -107,6 +107,16 @@ jobs:
           name: release-static
           path: build/bin/*
 
+      - name: Test offline Whisper
+        shell: bash
+        run: |
+          export PATH=$PWD/build/bin:$PATH
+          export EXE=sherpa-onnx-offline
+
+          readelf -d build/bin/sherpa-onnx-offline
+
+          .github/scripts/test-offline-whisper.sh
+
       - name: Test online CTC
         shell: bash
         run: |
@@ -139,16 +149,6 @@ jobs:
 
           .github/scripts/test-online-paraformer.sh
 
-      - name: Test offline Whisper
-        shell: bash
-        run: |
-          export PATH=$PWD/build/bin:$PATH
-          export EXE=sherpa-onnx-offline
-
-          readelf -d build/bin/sherpa-onnx-offline
-
-          .github/scripts/test-offline-whisper.sh
-
       - name: Test offline transducer
         shell: bash
         run: |

.github/workflows/windows-x86.yaml

@@ -93,13 +93,13 @@ jobs:
 
           .github/scripts/test-online-paraformer.sh
 
-      - name: Test offline Whisper for windows x86
-        shell: bash
-        run: |
-          export PATH=$PWD/build/bin/Release:$PATH
-          export EXE=sherpa-onnx-offline.exe
-
-          .github/scripts/test-offline-whisper.sh
+      # - name: Test offline Whisper for windows x86
+      #   shell: bash
+      #   run: |
+      #     export PATH=$PWD/build/bin/Release:$PATH
+      #     export EXE=sherpa-onnx-offline.exe
+      #
+      #     .github/scripts/test-offline-whisper.sh
 
       - name: Test offline CTC for windows x86
         shell: bash

scripts/whisper/export-onnx.py

@@ -8,6 +8,9 @@
 
 Thanks to https://github.com/TadaoYamaoka
 for making the onnx export script public.
+
+Note that we have removed the 30 seconds constraint from whisper. You can
+use any T <= 30.
 """
 
 import argparse
@@ -17,6 +20,7 @@
 
 import onnx
 import torch
+import torch.nn.functional as F
 from onnxruntime.quantization import QuantType, quantize_dynamic
 from torch import Tensor, nn
 
@@ -65,6 +69,39 @@ def add_meta_data(filename: str, meta_data: Dict[str, Any]):
     onnx.save(model, filename)
 
 
+def modified_audio_encoder_forward(self: AudioEncoder, x: torch.Tensor):
+    """
+    x : torch.Tensor, shape = (batch_size, n_mels, n_ctx)
+        the mel spectrogram of the audio
+    """
+    x = F.gelu(self.conv1(x))
+    x = F.gelu(self.conv2(x))
+    x = x.permute(0, 2, 1)
+
+    if False:
+        # This branch contains the original code
+        assert x.shape[1:] == self.positional_embedding.shape, "incorrect audio shape"
+        x = (x + self.positional_embedding).to(x.dtype)
+    else:
+        # This branch contains the actual changes
+        assert (
+            x.shape[2] == self.positional_embedding.shape[1]
+        ), f"incorrect audio shape: {x.shape}, {self.positional_embedding.shape}"
+        assert (
+            x.shape[1] == self.positional_embedding.shape[0]
+        ), f"incorrect audio shape: {x.shape}, {self.positional_embedding.shape}"
+        x = (x + self.positional_embedding[: x.shape[1]]).to(x.dtype)
+
+    for block in self.blocks:
+        x = block(x)
+
+    x = self.ln_post(x)
+    return x
+
+
+AudioEncoder.forward = modified_audio_encoder_forward
+
+
 class AudioEncoderTensorCache(nn.Module):
     def __init__(self, inAudioEncoder: AudioEncoder, inTextDecoder: TextDecoder):
         super().__init__()
@@ -279,6 +316,7 @@ def main():
         model = whisper.load_model(filename)
     else:
         model = whisper.load_model(name)
+    print(model.dims)
 
     print(
         f"number of model parameters: {name}",
@@ -311,19 +349,20 @@ def main():
     assert mel.shape == (batch_size, 80, 30 * 100)
 
     encoder = AudioEncoderTensorCache(model.encoder, model.decoder)
+
     n_layer_cross_k, n_layer_cross_v = encoder(mel)
     assert n_layer_cross_k.shape == (
         model.dims.n_text_layer,
         batch_size,
         model.dims.n_audio_ctx,
         model.dims.n_text_state,
-    ), n_layer_cross_k.shape
+    ), (n_layer_cross_k.shape, model.dims)
     assert n_layer_cross_v.shape == (
         model.dims.n_text_layer,
         batch_size,
         model.dims.n_audio_ctx,
         model.dims.n_text_state,
-    ), n_layer_cross_v.shape
+    ), (n_layer_cross_v.shape, model.dims)
 
     encoder_filename = f"{name}-encoder.onnx"
     torch.onnx.export(
@@ -334,9 +373,9 @@ def main():
         input_names=["mel"],
         output_names=["n_layer_cross_k", "n_layer_cross_v"],
         dynamic_axes={
-            "mel": {0: "n_audio"},  # n_audio is also known as batch_size
-            "n_layer_cross_k": {1: "n_audio"},
-            "n_layer_cross_v": {1: "n_audio"},
+            "mel": {0: "n_audio", 2: "T"},  # n_audio is also known as batch_size
+            "n_layer_cross_k": {1: "n_audio", 2: "T"},
+            "n_layer_cross_v": {1: "n_audio", 2: "T"},
         },
     )
 
@@ -461,8 +500,8 @@ def main():
             "tokens": {0: "n_audio", 1: "n_tokens"},
             "in_n_layer_self_k_cache": {1: "n_audio"},
             "in_n_layer_self_v_cache": {1: "n_audio"},
-            "n_layer_cross_k": {1: "n_audio"},
-            "n_layer_cross_v": {1: "n_audio"},
+            "n_layer_cross_k": {1: "n_audio", 2: "T"},
+            "n_layer_cross_v": {1: "n_audio", 2: "T"},
         },
     )
 

scripts/whisper/test.py

@@ -253,8 +253,21 @@ def compute_features(filename: str) -> torch.Tensor:
     log_spec = torch.clamp(features, min=1e-10).log10()
     log_spec = torch.maximum(log_spec, log_spec.max() - 8.0)
     mel = (log_spec + 4.0) / 4.0
+    # mel (T, 80)
+
+    # We pad 50 frames at the end so that it is able to detect eot
+    # You can use another value instead of 50.
+    mel = torch.nn.functional.pad(mel, (0, 0, 0, 50), "constant", 0)
+    # Note that if it throws for a multilingual model,
+    # please use a larger value, say 300
+
     target = 3000
-    mel = torch.nn.functional.pad(mel, (0, 0, 0, target - mel.shape[0]), "constant", 0)
+    if mel.shape[0] > target:
+        mel = mel[:target]
+
+    # We don't need to pad it to 30 seconds now!
+    #  mel = torch.nn.functional.pad(mel, (0, 0, 0, target - mel.shape[0]), "constant", 0)
+
     mel = mel.t().unsqueeze(0)
 
     return mel

sherpa-onnx/csrc/offline-recognizer-whisper-impl.h

@@ -115,15 +115,35 @@ class OfflineRecognizerWhisperImpl : public OfflineRecognizerImpl {
 
     NormalizeFeatures(f.data(), num_frames, feat_dim);
 
-    std::array<int64_t, 3> shape{1, max_num_frames, feat_dim};
+    // note that 50 is an experience value.
+    // see also ../../scripts/whisper/test.py
+    //
+    // You can replace 50 by other values, say, 100.
+    //
+    // Since we have removed the 30 seconds constraint, we need
+    // tail_padding_frames so that whisper is able to detect the eot token.
+    int32_t tail_padding_frames = 50;
+    if (model_->IsMultiLingual()) {
+      // 300 is an experience value. If it throws, please use a larger value.
+      tail_padding_frames = 300;
+    }
+
+    if (config_.model_config.whisper.tail_paddings > 0) {
+      tail_padding_frames = config_.model_config.whisper.tail_paddings;
+    }
+
+    int32_t actual_frames =
+        std::min(num_frames + tail_padding_frames, max_num_frames);
+
+    std::array<int64_t, 3> shape{1, actual_frames, feat_dim};
 
     Ort::Value mel = Ort::Value::CreateTensor<float>(
         model_->Allocator(), shape.data(), shape.size());
     float *p_mel = mel.GetTensorMutableData<float>();
     std::copy(f.begin(), f.end(), p_mel);
 
     memset(p_mel + f.size(), 0,
-           (max_num_frames - num_frames) * feat_dim * sizeof(float));
+           (actual_frames - num_frames) * feat_dim * sizeof(float));
     mel = Transpose12(model_->Allocator(), &mel);
 
     try {