Merge branch 'LASR-at-Home:main' into main

common/speech/lasr_speech_recognition_whisper/CMakeLists.txt

@@ -172,6 +172,7 @@ catkin_install_python(PROGRAMS
   scripts/test_microphones.py
   scripts/repeat_after_me.py
   scripts/test_speech_server.py
+  scripts/microphone_tuning_test.py
   DESTINATION ${CATKIN_PACKAGE_BIN_DESTINATION}
 )
 

common/speech/lasr_speech_recognition_whisper/scripts/microphone_tuning_test.py

@@ -0,0 +1,64 @@
+#!/usr/bin/env python3
+import argparse
+import os
+import torch
+import numpy as np
+from pathlib import Path
+import speech_recognition as sr
+from lasr_speech_recognition_whisper import load_model  # type: ignore
+import sounddevice  # needed to remove ALSA error messages
+from typing import Dict
+
+
+def parse_args() -> Dict:
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--device_index", type=int, default=None)
+    return vars(parser.parse_args())
+
+
+def configure_whisper_cache() -> None:
+    """Configures the whisper cache directory."""
+    whisper_cache = os.path.join(str(Path.home()), ".cache", "whisper")
+    os.makedirs(whisper_cache, exist_ok=True)
+    # Environemntal variable required to run whisper locally
+    os.environ["TIKTOKEN_CACHE_DIR"] = whisper_cache
+
+
+def main():
+    args = parse_args()
+
+    recognizer = sr.Recognizer()
+    microphone = sr.Microphone(device_index=args["device_index"], sample_rate=16000)
+    threshold = 100
+    recognizer.dynamic_energy_threshold = False
+    recognizer.energy_threshold = threshold
+    transcription_model = load_model(
+        "medium.en", "cuda" if torch.cuda.is_available() else "cpu", True
+    )
+    transcription_result = "The quick brown fox jumps over the lazy dog."
+    while transcription_result != "":
+        print(f"Listening...")
+        with microphone as source:
+            wav_data = recognizer.listen(source).get_wav_data()
+        print(f"Processing...")
+        # Magic number 32768.0 is the maximum value of a 16-bit signed integer
+        float_data = (
+            np.frombuffer(wav_data, dtype=np.int16).astype(np.float32, order="C")
+            / 32768.0
+        )
+
+        # Cast to fp16 if using GPU
+        transcription_result = transcription_model.transcribe(
+            float_data, fp16=torch.cuda.is_available()
+        )["text"]
+
+        print(
+            f"Transcription: {transcription_result} at energy threshold {recognizer.energy_threshold}"
+        )
+        threshold += 100
+        recognizer.energy_threshold = threshold
+
+
+if __name__ == "__main__":
+    configure_whisper_cache()
+    main()