chore: integrated local whisper + restructuring

.gitignore

@@ -1,3 +1,5 @@
+ggml-*.bin
+
 # Byte-compiled / optimized / DLL files
 __pycache__/
 *.py[cod]
@@ -7,7 +9,7 @@ __pycache__/
 *.so
 
 # Distribution / packaging
-.Python
+.Pythongit
 build/
 develop-eggs/
 dist/

OS/01/conversations/user.json

@@ -1 +1 @@
-[{"role": "user", "type": "message", "content": "This is a microphone. I also have an extra microphone.\n"}]
+[{"role": "user", "type": "message", "content": " Hello, how you doing?\n"}]

OS/01/local_stt/whisper-rust/README.md

@@ -0,0 +1,7 @@
+# Setup
+
+To rebuild the `whisper-rust` executable, do the following:
+
+1. Install [Rust](https://www.rust-lang.org/tools/install), cmake, and Python dependencies `pip install -r requirements.txt`.
+2. Go to **core/stt** and run `cargo build --release`.
+3. Move the `whisper-rust` executable from target/release to this directory.

OS/01/start.sh

@@ -1,7 +1,9 @@
 ### SETTINGS
 
 # If ALL_LOCAL is False, we'll use OpenAI's services
+# If setting ALL_LOCAL to true, set the path to the WHISPER local model
 export ALL_LOCAL=False
+# export WHISPER_MODEL_PATH=...
 # export OPENAI_API_KEY=sk-...
 
 # If SERVER_START, this is where we'll serve the server.

OS/01/stt.py

@@ -44,6 +44,28 @@ def export_audio_to_wav_ffmpeg(audio: bytearray, mime_type: str) -> str:
         os.remove(input_path)
         os.remove(output_path)
 
+def run_command(command):
+    result = subprocess.run(command, stdout=subprocess.PIPE, stderr=subprocess.PIPE, text=True)
+    return result.stdout, result.stderr
+
+def get_transcription_file(wav_file_path: str):
+    model_path = os.getenv("WHISPER_MODEL_PATH")
+    if not model_path:
+        raise EnvironmentError("WHISPER_MODEL_PATH environment variable is not set.")
+
+    output, error = run_command([
+        os.path.join(os.path.dirname(__file__), 'local_stt', 'whisper-rust', 'whisper-rust'),
+        '--model-path', model_path,
+        '--file-path', wav_file_path
+    ])
+
+    print("Exciting transcription result:", output)
+    return output
+
+def get_transcription_bytes(audio_bytes: bytearray, mime_type):
+    with export_audio_to_wav_ffmpeg(audio_bytes, mime_type) as wav_file_path:
+        return get_transcription_file(wav_file_path)
+
 def stt_bytes(audio_bytes: bytearray, mime_type="audio/wav"):
     with export_audio_to_wav_ffmpeg(audio_bytes, mime_type) as wav_file_path:
         return stt_wav(wav_file_path)
@@ -65,8 +87,12 @@ def stt_wav(wav_file_path: str):
         print("Transcription result:", transcript)
         return transcript
     else:
-        # Local whisper here, given `wav_file_path`
+        temp_dir = tempfile.gettempdir()
-        pass
+        output_path = os.path.join(temp_dir, f"output_{datetime.now().strftime('%Y%m%d%H%M%S%f')}.wav")
+        ffmpeg.input(wav_file_path).output(output_path, acodec='pcm_s16le', ac=1, ar='16k').run()
+        transcript = get_transcription_file(output_path)
+        print("Transcription result:", transcript)
+        return transcript
 
 def stt(input_data, mime_type="audio/wav"):
     if isinstance(input_data, str):

OS/_archive/core/stt/whisper-rust/.env.example

@@ -1 +0,0 @@
-WHISPER_MODEL_PATH=/path/to/ggml-tiny.en.bin

OS/_archive/core/stt/whisper-rust/README.md

@@ -1,9 +0,0 @@
-
-# Setup
-
-1. Install [Rust](https://www.rust-lang.org/tools/install) and Python dependencies `pip install -r requirements.txt`.
-2. Go to **core/stt** and run `cargo build --release`.
-3. Download GGML Whisper model from [Huggingface](https://huggingface.co/ggerganov/whisper.cpp).
-4. In core, copy `.env.example` to `.env` and put the path to model.
-5. Run `python core/i_endpoint.py` to start the server.
-6. Run `python core/test_cli.py PATH_TO_FILE` to test sending audio to service and getting transcription back over websocket.

OS/_archive/core/stt/whisper-rust/__init__.py

@@ -1,55 +0,0 @@
-from datetime import datetime
-import os
-import contextlib
-import tempfile
-import ffmpeg
-import subprocess
-
-def convert_mime_type_to_format(mime_type: str) -> str:
-    if mime_type == "audio/x-wav" or mime_type == "audio/wav":
-        return "wav"
-    if mime_type == "audio/webm":
-        return "webm"
-
-    return mime_type
-
-@contextlib.contextmanager
-def export_audio_to_wav_ffmpeg(audio: bytearray, mime_type: str) -> str:
-    temp_dir = tempfile.gettempdir()
-
-    # Create a temporary file with the appropriate extension
-    input_ext = convert_mime_type_to_format(mime_type)
-    input_path = os.path.join(temp_dir, f"input_{datetime.now().strftime('%Y%m%d%H%M%S%f')}.{input_ext}")
-    with open(input_path, 'wb') as f:
-        f.write(audio)
-
-    # Export to wav
-    output_path = os.path.join(temp_dir, f"output_{datetime.now().strftime('%Y%m%d%H%M%S%f')}.wav")
-    ffmpeg.input(input_path).output(output_path, acodec='pcm_s16le', ac=1, ar='16k').run()
-
-    print(f"Temporary file path: {output_path}")
-
-    try:
-        yield output_path
-    finally:
-        os.remove(input_path)
-        os.remove(output_path)
-
-def run_command(command):
-    result = subprocess.run(command, stdout=subprocess.PIPE, stderr=subprocess.PIPE, text=True)
-    return result.stdout, result.stderr
-
-def get_transcription(audio_bytes: bytearray, mime_type):
-    with export_audio_to_wav_ffmpeg(audio_bytes, mime_type) as wav_file_path:
-        model_path = os.getenv("WHISPER_MODEL_PATH")
-        if not model_path:
-            raise EnvironmentError("WHISPER_MODEL_PATH environment variable is not set.")
-
-        output, error = run_command([
-            os.path.join(os.path.dirname(__file__), 'whisper-rust', 'target', 'release', 'whisper-rust'),
-            '--model-path', model_path,
-            '--file-path', wav_file_path
-        ])
-
-        print("Exciting transcription result:", output)
-        return output

OS/_archive/core/stt/whisper-rust/test_cli.py

@@ -1,40 +0,0 @@
-import argparse
-import asyncio
-import websockets
-import os
-import json
-
-# Define the function to send audio file in chunks
-async def send_audio_in_chunks(file_path, chunk_size=4096):
-    async with websockets.connect("ws://localhost:8000/a") as websocket:
-        # Send the start command with mime type
-        await websocket.send(json.dumps({"action": "command", "state": "start", "mimeType": "audio/webm"}))
-
-        # Open the file in binary mode and send in chunks
-        with open(file_path, 'rb') as audio_file:
-            chunk = audio_file.read(chunk_size)
-            while chunk:
-                await websocket.send(chunk)
-                chunk = audio_file.read(chunk_size)
-
-        # Send the end command
-        await websocket.send(json.dumps({"action": "command", "state": "end"}))
-
-        # Receive a json message and then close the connection
-        message = await websocket.recv()
-        print("Received message:", json.loads(message))
-        await websocket.close()
-
-# Parse command line arguments
-parser = argparse.ArgumentParser(description="Send a webm audio file to the /a websocket endpoint and print the responses.")
-parser.add_argument("file_path", help="The path to the webm audio file to send.")
-args = parser.parse_args()
-
-# Check if the file exists
-if not os.path.isfile(args.file_path):
-    print(args.file_path)
-    print("Error: The file does not exist.")
-    exit(1)
-
-# Run the asyncio event loop
-asyncio.get_event_loop().run_until_complete(send_audio_in_chunks(args.file_path))

README.md

@@ -22,6 +22,10 @@ Official repository for [The 01 Project](https://twitter.com/hellokillian/status
    python -m pip install -r requirements.txt
    ```
 
+3. **(optional) Download local audio models**
+
+If you want to run local speech-to-text from whisper, download the GGML Whisper model from [Huggingface](https://huggingface.co/ggerganov/whisper.cpp). Then in `OS/01/start.sh`, set `ALL_LOCAL=TRUE` and set `WHISPER_MODEL_PATH` to the path of the model.
+
 ## Usage
 
 1. **Navigate to the project directory.**