[fix]: whisper_full_with_state: input is too short - 990 ms < 1000 ms. consider padding the input audio with silence.

transcribe/serve.py

@@ -92,7 +92,7 @@ class WhisperTranscriptionService:
                         self.full_segments_queue.appendleft(audio_array) # 根据时间是否满足三秒长度 来整合音频块
                         self.frames_np_start_timestamp = time.time()
                         self.frames_np = np.array([], dtype=np.float32)
-                    
+
                     # 音频结束信号的时候 整合当前缓冲区
                     #   START -- END -- START -- END 通常
                     #   START -- END -- END  end块带有音频信息的通常是4096内断的一个短音
@@ -105,7 +105,7 @@ class WhisperTranscriptionService:
                             self.frames_np = np.array([], dtype=np.float32)
                         else:
                             logger.debug(f"🥳 当前时间与上一句的时间差: {time_diff:.2f}s,继续保留在缓冲区")
-                    
+
             except queue.Empty:
                 pass
 
@@ -114,7 +114,7 @@ class WhisperTranscriptionService:
         frame_epoch = 1
 
         while not self._stop.is_set():
-      
+
             if len(self.frames_np) ==0:
                 time.sleep(0.1)
                 continue
@@ -126,11 +126,21 @@ class WhisperTranscriptionService:
                 else:
                     audio_buffer = self.frames_np[:int(frame_epoch * 1.5 * self.sample_rate)].copy()# 获取 1.5s * epoch 个音频长度
                     partial = True
-            
+
             if len(audio_buffer) < int(self.sample_rate):
-                silence_audio = np.zeros(self.sample_rate, dtype=np.float32)
-                silence_audio[-len(audio_buffer):] = audio_buffer
+                # Add a small buffer (e.g., 10ms worth of samples) to be safe
+                padding_samples = int(self.sample_rate * 0.01) # e.g., 160 samples for 10ms at 16kHz
+                target_length = self.sample_rate + padding_samples
+                silence_audio = np.zeros(target_length, dtype=np.float32)
+                # Ensure we don't try to copy more data than exists if audio_buffer is very short
+                copy_length = min(len(audio_buffer), target_length)
+                silence_audio[-copy_length:] = audio_buffer[-copy_length:] # Copy from the end of audio_buffer
                 audio_buffer = silence_audio
+            elif len(audio_buffer) > self.sample_rate * config.MAX_SPEECH_DURATION_S:
+                 # If buffer is too long even without padding, truncate it (optional, depends on desired behavior)
+                 # This case might already be handled elsewhere, but good to consider
+                 audio_buffer = audio_buffer[:int(self.sample_rate * config.MAX_SPEECH_DURATION_S)]
+
 
             logger.debug(f"audio buffer size: {len(audio_buffer) / self.sample_rate:.2f}s")
             meta_item = self._transcribe_audio(audio_buffer)
@@ -177,7 +187,7 @@ class WhisperTranscriptionService:
 
         result = self.translate_pipe.translate(text, self.source_language, self.target_language)
         translated_text = result.translate_content
- 
+
         log_block("🐧 Translation out ", f"{translated_text}")
         return translated_text