fix a bug for wenet streaming model. (#1054)

csukuangfj · web-flow · commit a3bac19c5480 · 2024-06-24T21:52:54.000+08:00
* fix a bug for wenet streaming model. The chunk shift was wrong. See https://github.com/wenet-e2e/wenet/blob/main/runtime/core/decoder/asr_model.cc#L15 and https://github.com/wenet-e2e/wenet/blob/main/runtime/core/decoder/asr_model.cc#L28
diff --git a/.github/workflows/test-build-wheel.yaml b/.github/workflows/test-build-wheel.yaml
@@ -132,7 +132,7 @@ jobs:
           export PATH=/c/hostedtoolcache/windows/Python/3.9.13/x64/bin:$PATH
           export PATH=/c/hostedtoolcache/windows/Python/3.10.11/x64/bin:$PATH
           export PATH=/c/hostedtoolcache/windows/Python/3.11.9/x64/bin:$PATH
-          export PATH=/c/hostedtoolcache/windows/Python/3.12.3/x64/bin:$PATH
+          export PATH=/c/hostedtoolcache/windows/Python/3.12.4/x64/bin:$PATH
 
           which sherpa-onnx
           sherpa-onnx --help
diff --git a/.github/workflows/test-pip-install.yaml b/.github/workflows/test-pip-install.yaml
@@ -107,7 +107,7 @@ jobs:
           export PATH=/c/hostedtoolcache/windows/Python/3.9.13/x64/bin:$PATH
           export PATH=/c/hostedtoolcache/windows/Python/3.10.11/x64/bin:$PATH
           export PATH=/c/hostedtoolcache/windows/Python/3.11.9/x64/bin:$PATH
-          export PATH=/c/hostedtoolcache/windows/Python/3.12.3/x64/bin:$PATH
+          export PATH=/c/hostedtoolcache/windows/Python/3.12.4/x64/bin:$PATH
 
           sherpa-onnx --help
           sherpa-onnx-keyword-spotter --help
diff --git a/scripts/wenet/test-onnx-streaming.py b/scripts/wenet/test-onnx-streaming.py
@@ -143,7 +143,7 @@ def main():
         (model.chunk_size - 1) * model.subsampling_factor + model.right_context + 1
     )
     chunk_length = int(chunk_length)
-    chunk_shift = int(model.required_cache_size)
+    chunk_shift = int(model.chunk_size * model.subsampling_factor)
     print(chunk_length, chunk_shift)
 
     num_frames = x.shape[0]
diff --git a/sherpa-onnx/csrc/online-wenet-ctc-model.cc b/sherpa-onnx/csrc/online-wenet-ctc-model.cc
@@ -97,7 +97,9 @@ class OnlineWenetCtcModel::Impl {
            right_context_ + 1;
   }
 
-  int32_t ChunkShift() const { return required_cache_size_; }
+  int32_t ChunkShift() const {
+    return config_.wenet_ctc.chunk_size * subsampling_factor_;
+  }
 
   OrtAllocator *Allocator() const { return allocator_; }
 

Original file line number	Diff line number	Diff line change
`@@ -143,7 +143,7 @@ def main():`
`143`	`143`	`(model.chunk_size - 1) * model.subsampling_factor + model.right_context + 1`
`144`	`144`	`)`
`145`	`145`	`chunk_length = int(chunk_length)`
`146`		`- chunk_shift = int(model.required_cache_size)`
	`146`	`+ chunk_shift = int(model.chunk_size * model.subsampling_factor)`
`147`	`147`	`print(chunk_length, chunk_shift)`
`148`	`148`
`149`	`149`	`num_frames = x.shape[0]`
Original file line number	Diff line number	Diff line change
`@@ -97,7 +97,9 @@ class OnlineWenetCtcModel::Impl {`
`97`	`97`	`right_context_ + 1;`
`98`	`98`	`}`
`99`	`99`
`100`		`- int32_t ChunkShift() const { return required_cache_size_; }`
	`100`	`+ int32_t ChunkShift() const {`
	`101`	`+ return config_.wenet_ctc.chunk_size * subsampling_factor_;`
	`102`	`+ }`
`101`	`103`
`102`	`104`	`OrtAllocator *Allocator() const { return allocator_; }`
`103`	`105`