streaming_decode.py, relax the audio range from [-1,+1] to [-10,+10] (#1448)

- some AudioTransform classes produce audio signals out of range [-1,+1]
   - Resample produced 1.0079
   - The range [-10,+10] was chosen to still be able to reliably
     distinguish from the [-32k,+32k] signal...
- this is related to : https://github.com/lhotse-speech/lhotse/issues/1254

egs/aishell/ASR/pruned_transducer_stateless7_streaming/streaming_decode.py

@@ -342,7 +342,12 @@ def decode_dataset(
         assert audio.dtype == np.float32, audio.dtype
 
         # The trained model is using normalized samples
-        assert audio.max() <= 1, "Should be normalized to [-1, 1])"
+        # - this is to avoid sending [-32k,+32k] signal in...
+        # - some lhotse AudioTransform classes can make the signal
+        #   be out of range [-1, 1], hence the tolerance 10
+        assert (
+            np.abs(audio).max() <= 10
+        ), "Should be normalized to [-1, 1], 10 for tolerance..."
 
         samples = torch.from_numpy(audio).squeeze(0)
 

egs/aishell/ASR/zipformer/streaming_decode.py

@@ -597,12 +597,12 @@ def decode_dataset(
         assert audio.dtype == np.float32, audio.dtype
 
         # The trained model is using normalized samples
-        if audio.max() > 1:
+        # - this is to avoid sending [-32k,+32k] signal in...
-            logging.warning(
+        # - some lhotse AudioTransform classes can make the signal
-                f"The audio should be normalized to [-1, 1], audio.max : {audio.max()}."
+        #   be out of range [-1, 1], hence the tolerance 10
-                f"Clipping to [-1, 1]."
+        assert (
-            )
+            np.abs(audio).max() <= 10
-            audio = np.clip(audio, -1, 1)
+        ), "Should be normalized to [-1, 1], 10 for tolerance..."
 
         samples = torch.from_numpy(audio).squeeze(0)
 

egs/csj/ASR/pruned_transducer_stateless7_streaming/streaming_decode.py

@@ -362,7 +362,12 @@ def decode_dataset(
         assert audio.dtype == np.float32, audio.dtype
 
         # The trained model is using normalized samples
-        assert audio.max() <= 1, "Should be normalized to [-1, 1])"
+        # - this is to avoid sending [-32k,+32k] signal in...
+        # - some lhotse AudioTransform classes can make the signal
+        #   be out of range [-1, 1], hence the tolerance 10
+        assert (
+            np.abs(audio).max() <= 10
+        ), "Should be normalized to [-1, 1], 10 for tolerance..."
 
         samples = torch.from_numpy(audio).squeeze(0)
 

egs/gigaspeech/ASR/zipformer/streaming_decode.py

@@ -578,7 +578,12 @@ def decode_dataset(
         assert audio.dtype == np.float32, audio.dtype
 
         # The trained model is using normalized samples
-        assert audio.max() <= 1, "Should be normalized to [-1, 1])"
+        # - this is to avoid sending [-32k,+32k] signal in...
+        # - some lhotse AudioTransform classes can make the signal
+        #   be out of range [-1, 1], hence the tolerance 10
+        assert (
+            np.abs(audio).max() <= 10
+        ), "Should be normalized to [-1, 1], 10 for tolerance..."
 
         samples = torch.from_numpy(audio).squeeze(0)
 

egs/librispeech/ASR/conv_emformer_transducer_stateless/streaming_decode.py

@@ -681,8 +681,14 @@ def decode_dataset(
         assert len(audio.shape) == 2
         assert audio.shape[0] == 1, "Should be single channel"
         assert audio.dtype == np.float32, audio.dtype
+
         # The trained model is using normalized samples
-        assert audio.max() <= 1, "Should be normalized to [-1, 1])"
+        # - this is to avoid sending [-32k,+32k] signal in...
+        # - some lhotse AudioTransform classes can make the signal
+        #   be out of range [-1, 1], hence the tolerance 10
+        assert (
+            np.abs(audio).max() <= 10
+        ), "Should be normalized to [-1, 1], 10 for tolerance..."
 
         samples = torch.from_numpy(audio).squeeze(0)
         feature = fbank(samples)

egs/librispeech/ASR/conv_emformer_transducer_stateless2/streaming_decode.py

@@ -681,8 +681,14 @@ def decode_dataset(
         assert len(audio.shape) == 2
         assert audio.shape[0] == 1, "Should be single channel"
         assert audio.dtype == np.float32, audio.dtype
+
         # The trained model is using normalized samples
-        assert audio.max() <= 1, "Should be normalized to [-1, 1])"
+        # - this is to avoid sending [-32k,+32k] signal in...
+        # - some lhotse AudioTransform classes can make the signal
+        #   be out of range [-1, 1], hence the tolerance 10
+        assert (
+            np.abs(audio).max() <= 10
+        ), "Should be normalized to [-1, 1], 10 for tolerance..."
 
         samples = torch.from_numpy(audio).squeeze(0)
         feature = fbank(samples)

egs/librispeech/ASR/lstm_transducer_stateless/streaming_decode.py

@@ -673,8 +673,14 @@ def decode_dataset(
         assert len(audio.shape) == 2
         assert audio.shape[0] == 1, "Should be single channel"
         assert audio.dtype == np.float32, audio.dtype
+
         # The trained model is using normalized samples
-        assert audio.max() <= 1, "Should be normalized to [-1, 1])"
+        # - this is to avoid sending [-32k,+32k] signal in...
+        # - some lhotse AudioTransform classes can make the signal
+        #   be out of range [-1, 1], hence the tolerance 10
+        assert (
+            np.abs(audio).max() <= 10
+        ), "Should be normalized to [-1, 1], 10 for tolerance..."
 
         samples = torch.from_numpy(audio).squeeze(0)
         feature = fbank(samples)

egs/librispeech/ASR/lstm_transducer_stateless3/streaming_decode.py

@@ -673,8 +673,14 @@ def decode_dataset(
         assert len(audio.shape) == 2
         assert audio.shape[0] == 1, "Should be single channel"
         assert audio.dtype == np.float32, audio.dtype
+
         # The trained model is using normalized samples
-        assert audio.max() <= 1, "Should be normalized to [-1, 1])"
+        # - this is to avoid sending [-32k,+32k] signal in...
+        # - some lhotse AudioTransform classes can make the signal
+        #   be out of range [-1, 1], hence the tolerance 10
+        assert (
+            np.abs(audio).max() <= 10
+        ), "Should be normalized to [-1, 1], 10 for tolerance..."
 
         samples = torch.from_numpy(audio).squeeze(0)
         feature = fbank(samples)

egs/librispeech/ASR/pruned_transducer_stateless/streaming_decode.py

@@ -359,7 +359,12 @@ def decode_dataset(
         assert audio.dtype == np.float32, audio.dtype
 
         # The trained model is using normalized samples
-        assert audio.max() <= 1, "Should be normalized to [-1, 1])"
+        # - this is to avoid sending [-32k,+32k] signal in...
+        # - some lhotse AudioTransform classes can make the signal
+        #   be out of range [-1, 1], hence the tolerance 10
+        assert (
+            np.abs(audio).max() <= 10
+        ), "Should be normalized to [-1, 1], 10 for tolerance..."
 
         samples = torch.from_numpy(audio).squeeze(0)
 

egs/librispeech/ASR/pruned_transducer_stateless2/streaming_decode.py

@@ -361,7 +361,12 @@ def decode_dataset(
         assert audio.dtype == np.float32, audio.dtype
 
         # The trained model is using normalized samples
-        assert audio.max() <= 1, "Should be normalized to [-1, 1])"
+        # - this is to avoid sending [-32k,+32k] signal in...
+        # - some lhotse AudioTransform classes can make the signal
+        #   be out of range [-1, 1], hence the tolerance 10
+        assert (
+            np.abs(audio).max() <= 10
+        ), "Should be normalized to [-1, 1], 10 for tolerance..."
 
         samples = torch.from_numpy(audio).squeeze(0)
 

egs/librispeech/ASR/pruned_transducer_stateless3/streaming_decode.py

@@ -362,7 +362,12 @@ def decode_dataset(
         assert audio.dtype == np.float32, audio.dtype
 
         # The trained model is using normalized samples
-        assert audio.max() <= 1, "Should be normalized to [-1, 1])"
+        # - this is to avoid sending [-32k,+32k] signal in...
+        # - some lhotse AudioTransform classes can make the signal
+        #   be out of range [-1, 1], hence the tolerance 10
+        assert (
+            np.abs(audio).max() <= 10
+        ), "Should be normalized to [-1, 1], 10 for tolerance..."
 
         samples = torch.from_numpy(audio).squeeze(0)
 

egs/librispeech/ASR/pruned_transducer_stateless4/streaming_decode.py

@@ -378,7 +378,12 @@ def decode_dataset(
         assert audio.dtype == np.float32, audio.dtype
 
         # The trained model is using normalized samples
-        assert audio.max() <= 1, "Should be normalized to [-1, 1])"
+        # - this is to avoid sending [-32k,+32k] signal in...
+        # - some lhotse AudioTransform classes can make the signal
+        #   be out of range [-1, 1], hence the tolerance 10
+        assert (
+            np.abs(audio).max() <= 10
+        ), "Should be normalized to [-1, 1], 10 for tolerance..."
 
         samples = torch.from_numpy(audio).squeeze(0)
 

egs/librispeech/ASR/pruned_transducer_stateless5/streaming_decode.py

@@ -378,7 +378,12 @@ def decode_dataset(
         assert audio.dtype == np.float32, audio.dtype
 
         # The trained model is using normalized samples
-        assert audio.max() <= 1, "Should be normalized to [-1, 1])"
+        # - this is to avoid sending [-32k,+32k] signal in...
+        # - some lhotse AudioTransform classes can make the signal
+        #   be out of range [-1, 1], hence the tolerance 10
+        assert (
+            np.abs(audio).max() <= 10
+        ), "Should be normalized to [-1, 1], 10 for tolerance..."
 
         samples = torch.from_numpy(audio).squeeze(0)
 

egs/librispeech/ASR/pruned_transducer_stateless7_streaming/streaming_decode.py

@@ -345,7 +345,12 @@ def decode_dataset(
         assert audio.dtype == np.float32, audio.dtype
 
         # The trained model is using normalized samples
-        assert audio.max() <= 1, "Should be normalized to [-1, 1])"
+        # - this is to avoid sending [-32k,+32k] signal in...
+        # - some lhotse AudioTransform classes can make the signal
+        #   be out of range [-1, 1], hence the tolerance 10
+        assert (
+            np.abs(audio).max() <= 10
+        ), "Should be normalized to [-1, 1], 10 for tolerance..."
 
         samples = torch.from_numpy(audio).squeeze(0)
 

egs/librispeech/ASR/pruned_transducer_stateless7_streaming_multi/streaming_decode.py

@@ -345,7 +345,12 @@ def decode_dataset(
         assert audio.dtype == np.float32, audio.dtype
 
         # The trained model is using normalized samples
-        assert audio.max() <= 1, "Should be normalized to [-1, 1])"
+        # - this is to avoid sending [-32k,+32k] signal in...
+        # - some lhotse AudioTransform classes can make the signal
+        #   be out of range [-1, 1], hence the tolerance 10
+        assert (
+            np.abs(audio).max() <= 10
+        ), "Should be normalized to [-1, 1], 10 for tolerance..."
 
         samples = torch.from_numpy(audio).squeeze(0)
 

egs/librispeech/ASR/zipformer/streaming_decode.py

@@ -577,7 +577,12 @@ def decode_dataset(
         assert audio.dtype == np.float32, audio.dtype
 
         # The trained model is using normalized samples
-        assert audio.max() <= 1, "Should be normalized to [-1, 1])"
+        # - this is to avoid sending [-32k,+32k] signal in...
+        # - some lhotse AudioTransform classes can make the signal
+        #   be out of range [-1, 1], hence the tolerance 10
+        assert (
+            np.abs(audio).max() <= 10
+        ), "Should be normalized to [-1, 1], 10 for tolerance..."
 
         samples = torch.from_numpy(audio).squeeze(0)
 

egs/wenetspeech/ASR/pruned_transducer_stateless5/streaming_decode.py

@@ -402,6 +402,14 @@ def decode_dataset(
         assert audio.shape[0] == 1, "Should be single channel"
         assert audio.dtype == np.float32, audio.dtype
 
+        # The trained model is using normalized samples
+        # - this is to avoid sending [-32k,+32k] signal in...
+        # - some lhotse AudioTransform classes can make the signal
+        #   be out of range [-1, 1], hence the tolerance 10
+        assert (
+            np.abs(audio).max() <= 10
+        ), "Should be normalized to [-1, 1], 10 for tolerance..."
+
         samples = torch.from_numpy(audio).squeeze(0)
 
         fbank = Fbank(opts)

egs/wenetspeech/ASR/zipformer/streaming_decode.py

@@ -597,12 +597,12 @@ def decode_dataset(
         assert audio.dtype == np.float32, audio.dtype
 
         # The trained model is using normalized samples
-        if audio.max() > 1:
+        # - this is to avoid sending [-32k,+32k] signal in...
-            logging.warning(
+        # - some lhotse AudioTransform classes can make the signal
-                f"The audio should be normalized to [-1, 1], audio.max : {audio.max()}."
+        #   be out of range [-1, 1], hence the tolerance 10
-                f"Clipping to [-1, 1]."
+        assert (
-            )
+            np.abs(audio).max() <= 10
-            audio = np.clip(audio, -1, 1)
+        ), "Should be normalized to [-1, 1], 10 for tolerance..."
 
         samples = torch.from_numpy(audio).squeeze(0)