Merge pull request #43 from duaneking/master

Really make it real time with no disk IO needed.

requirements.txt

@@ -2,4 +2,5 @@ pyaudio
 SpeechRecognition
 --extra-index-url https://download.pytorch.org/whl/cu116
 torch
+numpy
 git+https://github.com/openai/whisper.git

transcribe_demo.py

@@ -1,15 +1,14 @@
 #! python3.7
 
 import argparse
-import io
 import os
+import numpy as np
 import speech_recognition as sr
 import whisper
 import torch
 
 from datetime import datetime, timedelta
 from queue import Queue
-from tempfile import NamedTemporaryFile
 from time import sleep
 from sys import platform
 
@@ -35,8 +34,6 @@ def main():
 
     # The last time a recording was retrieved from the queue.
     phrase_time = None
-    # Current raw audio bytes.
-    last_sample = bytes()
     # Thread safe Queue for passing data from the threaded recording callback.
     data_queue = Queue()
     # We use SpeechRecognizer to record our audio because it has a nice feature where it can detect when speech ends.
@@ -71,7 +68,6 @@ def main():
     record_timeout = args.record_timeout
     phrase_timeout = args.phrase_timeout
 
-    temp_file = NamedTemporaryFile().name
     transcription = ['']
 
     with source:
@@ -102,26 +98,21 @@ def main():
                 # If enough time has passed between recordings, consider the phrase complete.
                 # Clear the current working audio buffer to start over with the new data.
                 if phrase_time and now - phrase_time > timedelta(seconds=phrase_timeout):
-                    last_sample = bytes()
                     phrase_complete = True
                 # This is the last time we received new audio data from the queue.
                 phrase_time = now
                 
-                # Concatenate our current audio data with the latest audio data.
-                while not data_queue.empty():
-                    data = data_queue.get()
-                    last_sample += data
+                # Combine audio data from queue
+                audio_data = b''.join(data_queue.queue)
+                data_queue.queue.clear()
                 
-                # Use AudioData to convert the raw data to wav data.
-                audio_data = sr.AudioData(last_sample, source.SAMPLE_RATE, source.SAMPLE_WIDTH)
-                wav_data = io.BytesIO(audio_data.get_wav_data())
-
-                # Write wav data to the temporary file as bytes.
-                with open(temp_file, 'w+b') as f:
-                    f.write(wav_data.read())
+                # Convert in-ram buffer to something the model can use directly without needing a temp file.
+                # Convert data from 16 bit wide integers to floating point with a width of 32 bits.
+                # Clamp the audio stream frequency to a PCM wavelength compatible default of 32768hz max.
+                audio_np = np.frombuffer(audio_data, dtype=np.int16).astype(np.float32) / 32768.0
 
                 # Read the transcription.
-                result = audio_model.transcribe(temp_file, fp16=torch.cuda.is_available())
+                result = audio_model.transcribe(audio_np, fp16=torch.cuda.is_available())
                 text = result['text'].strip()
 
                 # If we detected a pause between recordings, add a new item to our transcription.