Pad 1s of silence on endpoint for streaming paraformer in Python examples (mirror C++ behavior, #1373) (#4052)

python-api-examples/streaming-paraformer-asr-microphone.py

@@ -9,6 +9,8 @@
 # to download pre-trained models
 
 import sys
+
+import numpy as np
 from pathlib import Path
 
 try:
@@ -92,6 +94,20 @@ def main():
             display.display()
 
             if is_endpoint:
+                # For streaming paraformer models, which have a large right
+                # chunk size, we need to pad ~1 second of silence on
+                # endpointing so that the last character can be recognized.
+                # (Mirrors the C++ example sherpa-onnx-microphone.cc; see
+                # https://github.com/k2-fsa/sherpa-onnx/issues/1373)
+                tail_paddings = np.zeros(int(1.0 * sample_rate), dtype=np.float32)
+                stream.accept_waveform(sample_rate, tail_paddings)
+                while recognizer.is_ready(stream):
+                    recognizer.decode_stream(stream)
+
+                result = recognizer.get_result(stream)
+                display.update_text(result)
+                display.display()
+
                 if result:
                     display.finalize_current_sentence()
                     display.display()

python-api-examples/streaming_server.py

@@ -757,6 +757,26 @@ Go back to <a href="/streaming_record.html">/streaming_record.html</a>
                     "segment": segment,
                 }
                 if self.recognizer.is_endpoint(stream):
+                    if self.recognizer.config.model_config.paraformer.encoder:
+                        # For streaming paraformer models, which have a large
+                        # right chunk size, we need to pad ~1 second of silence
+                        # on endpointing so that the last character can be
+                        # recognized. (Mirrors sherpa-onnx-microphone.cc; see
+                        # https://github.com/k2-fsa/sherpa-onnx/issues/1373)
+                        tail_padding = np.zeros(self.sample_rate).astype(np.float32)
+                        stream.accept_waveform(
+                            sample_rate=self.sample_rate, waveform=tail_padding
+                        )
+                        while self.recognizer.is_ready(stream):
+                            await self.compute_and_decode(stream)
+
+                        result = self.recognizer.get_result(stream)
+                        message = {
+                            "text": result,
+                            "segment": segment,
+                        }
+                        await socket.send(json.dumps(message))
+
                     self.recognizer.reset(stream)
                     segment += 1