Whisper streaming support for CPU

comps/asr/src/requirements.txt

@@ -3,6 +3,7 @@ aiohttp
 datasets
 docarray[full]
 fastapi
+faster-whisper
 mcp
 opentelemetry-api
 opentelemetry-exporter-otlp

comps/third_parties/whisper/src/faster_whisper_model.py

@@ -0,0 +1,97 @@
+# Copyright (C) 2025 Intel Corporation
+# SPDX-License-Identifier: Apache-2.0
+
+import os
+import uuid
+import wave
+
+from fastapi import WebSocket
+from faster_whisper import WhisperModel
+
+BYTES_PER_SAMPLE = 2  # 16-bit audio
+SAMPLE_RATE = 16000  # Hz
+
+
+class FasterWhisperModel:
+    """Leverage FasterWhisper to do realtime transcription on CPU."""
+
+    def __init__(self, model_size_or_path="tiny", device="cpu", compute_type="int8") -> None:
+        self.model = WhisperModel(
+            model_size_or_path=model_size_or_path, device=device, compute_type=compute_type, local_files_only=False
+        )
+        self.model_size_or_path = model_size_or_path
+
+    async def audio2text_streaming(
+        self,
+        websocket: WebSocket,
+        audio_data: bytes,
+        item_id: int,
+        event_id: str,
+        vad_filter: bool = True,
+        beam_size: int = 5,
+        language: str = "en",
+        min_silence_duration_ms: int = 500,
+        is_final: bool = False,
+    ):
+        """Process the audio data chunk."""
+        if not audio_data:
+            return
+        # initialize the transcription variable to concatenate the transcription
+        transcription = ""
+
+        # generate a temporary file name
+        uid = str(uuid.uuid4())
+        temp_filename = f"{uid}.wav"
+
+        try:
+            # create a WAV file
+            with wave.open(temp_filename, "wb") as wav_file:
+                wav_file.setnchannels(1)  # single channel
+                wav_file.setsampwidth(BYTES_PER_SAMPLE)  # 16-bit
+                wav_file.setframerate(SAMPLE_RATE)
+                wav_file.writeframes(audio_data)
+
+            # use faster-whisper for speech recognition
+            segments, _ = self.model.transcribe(
+                temp_filename,
+                language=None if language == "auto" else language,
+                beam_size=beam_size,
+                vad_filter=vad_filter,
+                vad_parameters=dict(min_silence_duration_ms=min_silence_duration_ms),
+            )
+
+            segments = list(segments)
+            i = 0
+            for segment in segments:
+                print("[%.2fs -> %.2fs] %s" % (segment.start, segment.end, segment.text))
+                response = {
+                    "content_index": i,
+                    "type": "conversation.item.input_audio_transcription.delta",
+                    "event_id": event_id,
+                    "item_id": "item_" + f"{item_id:03}",
+                    "delta": segment.text,
+                }
+                transcription += segment.text
+                await websocket.send_json(response)
+
+            # if it is the last chunk of data, send the completion signal
+            if is_final:
+                # Send completion signal for streaming format
+                await websocket.send_json(
+                    {
+                        "content_index": 0,
+                        "type": "conversation.item.input_audio_transcription.completion",
+                        "event_id": event_id,
+                        "item_id": "item_" + f"{item_id:03}",
+                        "transcript": transcription,
+                    }
+                )
+                transcription = ""
+
+        except Exception as e:
+            print(f"Error in transcription: {e}")
+            await websocket.send_json({"status": "error", "error": str(e)})
+        finally:
+            # clean up the temporary file
+            if os.path.exists(temp_filename):
+                os.remove(temp_filename)