From 74afc2f9f0db3e589ad9b4541ed7d2685eed2344 Mon Sep 17 00:00:00 2001
From: Mathieu Broillet <mathieu@broillet.ch>
Date: Wed, 31 May 2023 17:16:09 +0200
Subject: [PATCH] replaced whispercpp by faster-whisper

---
 jarvis/api.py                        |  6 +++--
 jarvis/utils/faster_whisper_utils.py | 37 ++++++++++++++++++++++++++++
 start.py                             | 11 ++++-----
 3 files changed, 46 insertions(+), 8 deletions(-)
 create mode 100644 jarvis/utils/faster_whisper_utils.py

diff --git a/jarvis/api.py b/jarvis/api.py
index 227a83c..c5a8205 100644
--- a/jarvis/api.py
+++ b/jarvis/api.py
@@ -9,7 +9,7 @@ from flask import Flask, request
 from flask_socketio import SocketIO, emit, join_room, leave_room, \
     rooms
 
-from jarvis.utils import chat_utils, whisper_utils, chatgpt_utils
+from jarvis.utils import chat_utils, whisper_utils, chatgpt_utils, faster_whisper_utils
 
 # Set this variable to "threading", "eventlet" or "gevent" to test the
 # different async modes, or leave it set to None for the application to choose
@@ -86,7 +86,9 @@ def get_text_from_audio():
     audio_temp_file = tempfile.NamedTemporaryFile(prefix='jarvis-audio_', suffix='_client')
     audio_temp_file.write(request.data)
 
-    text = whisper_utils.whisper_cpp_stt(audio_temp_file.name)
+    # text = whisper_utils.whisper_cpp_stt(audio_temp_file.name)
+    text = faster_whisper_utils.faster_whisper_stt(audio_temp_file.name)
+
     logging.info("STT result for " + request.remote_addr + " : " + text)
 
     return {"data": text}
diff --git a/jarvis/utils/faster_whisper_utils.py b/jarvis/utils/faster_whisper_utils.py
new file mode 100644
index 0000000..15a93a0
--- /dev/null
+++ b/jarvis/utils/faster_whisper_utils.py
@@ -0,0 +1,37 @@
+import logging
+
+from faster_whisper import WhisperModel
+
+
+def load_model():
+    log_level = logging.getLogger().level
+    global model
+    model = WhisperModel('small', device="cpu", cpu_threads=8, compute_type="int8")
+    logging.getLogger().setLevel(log_level)
+
+
+def get_model():
+    return model
+
+
+def faster_whisper_stt(audio_file):
+    """
+    Transcribe audio file using whisper-cpp, no additional server/service needed, runs on CPU.
+
+    :param audio_file:
+    :param model:
+    :return: text
+    """
+    if model is None:
+        logging.error("Model is not loaded")
+        load_model()
+
+    segments, info = model.transcribe(audio_file, beam_size=5)
+    print("Detected language '%s' with probability %f" % (info.language, info.language_probability))
+
+    # combines all segments in one string
+    text = ''
+    for segment in segments:
+        text += segment.text + ' '
+
+    return text
diff --git a/start.py b/start.py
index 9ced225..a201c52 100644
--- a/start.py
+++ b/start.py
@@ -3,9 +3,7 @@ import logging
 import lingua_franca
 
 import jarvis.api
-from jarvis.skills.cocktails import CocktailSkill
-from jarvis.skills.intent_services import intent_manager
-from jarvis.utils import whisper_utils
+from jarvis.utils import faster_whisper_utils
 
 if __name__ == '__main__':
     logging.getLogger().setLevel(logging.INFO)
@@ -14,13 +12,14 @@ if __name__ == '__main__':
     lingua_franca.load_language(lang="fr")
 
     # Register each skills
-    CocktailSkill().register()
+    # CocktailSkill().register()
 
     # Load the skills
-    intent_manager.load_all_skills()
+    # intent_manager.load_all_skills()
 
     # Load the STT (whisper) model
-    whisper_utils.load_model()
+    # whisper_utils.load_model()
+    faster_whisper_utils.load_model()
 
     # Start the api endpoint
     jarvis.api.start_api()