Add option to reduce bandwidth (and potentially latency) by converting
voice recordings (wav) into a compressed audio format (webm or mp3).
Default behaviour is unchanged.
> File uploads are currently limited to 25 MB and the following input file
> types are supported: mp3, mp4, mpeg, mpga, m4a, wav, and webm.
>
> - https://platform.openai.com/docs/guides/speech-to-text
# Aider chat conversation:
USER: what is `indata` ?
ASSISTANT: `indata` is a parameter in the callback function defined within the `record_and_transcribe` function in the `aider/voice.py` file. This parameter represents the input audio data that is being recorded.
In the context of the `sounddevice` library, which is used for audio recording in this script, `indata` is a 2D numpy array containing the recorded audio data. The rows in the array represent individual frames, while the columns represent channels.
In this specific script, the audio is being recorded in mono (single channel), so there will be only one column. The data in `indata` is being put into a queue and also being dumped for debugging purposes.
USER: how can i estimate the avg volume based on indata?
ASSISTANT: To estimate the average volume of the audio data, you can calculate the root mean square (RMS) of the `indata` array. The RMS represents the square root of the average of the squares of the values, and it's a common way to measure the magnitude of a varying quantity, such as audio signals.
Here's how you can modify the `record_and_transcribe` function to calculate the RMS:
```python
aider/voice.py
<<<<<<< HEAD
def callback(indata, frames, time, status):
"""This is called (from a separate thread) for each audio block."""
q.put(indata.copy())
dump(indata)
dump(indata.mean())
=======
import numpy as np
def callback(indata, frames, time, status):
"""This is called (from a separate thread) for each audio block."""
q.put(indata.copy())
rms = np.sqrt(np.mean(indata**2))
dump(rms)
>>>>>>> updated
```
In this change, we import the numpy library and replace the line `dump(indata.mean())` with `rms = np.sqrt(np.mean(indata**2))` and `dump(rms)`. This calculates the RMS of the `indata` and dumps it for debugging purposes.
# Aider chat conversation:
USER: I ran this command:
python -m aider.voice
And got this output:
Traceback (most recent call last):
File "<frozen runpy>", line 198, in _run_module_as_main
File "<frozen runpy>", line 88, in _run_code
File "/Users/gauthier/Projects/aider/aider/voice.py", line 42, in <module>
print(record_and_transcribe(api_key))
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/Users/gauthier/Projects/aider/aider/voice.py", line 22, in record_and_transcribe
recording = sd.rec(int(sample_rate * duration), samplerate=sample_rate, channels=1, callback=on_key_press)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/Users/gauthier/Projects/aider/.venv/lib/python3.11/site-packages/sounddevice.py", line 276, in rec
ctx.start_stream(InputStream, samplerate, ctx.input_channels,
TypeError: _CallbackContext.start_stream() got multiple values for argument 'callback'
Recording started, press any key to stop...
ASSISTANT: It seems like there's an issue with the way the `callback` argument is being passed to the `sd.rec` function in the `record_and_transcribe` function in the `aider/voice.py` file.
The `sd.rec` function is a convenience function that internally creates an instance of `sd.InputStream` and starts recording. The `callback` argument should be passed to the `sd.InputStream` constructor, not the `sd.rec` function.
Here's how we can fix this:
1. Modify the `aider/voice.py` file.
```python
aider/voice.py
<<<<<<< HEAD
recording = sd.rec(int(sample_rate * duration), samplerate=sample_rate, channels=1, callback=on_key_press)
=======
# Create an instance of InputStream with the callback
stream = sd.InputStream(samplerate=sample_rate, channels=1, callback=on_key_press)
stream.start()
recording = sd.rec(int(sample_rate * duration), samplerate=sample_rate, channels=1)
>>>>>>> updated
```
This change creates an instance of `sd.InputStream` with the `on_key_press` callback and starts the stream before calling `sd.rec`. This should ensure that the callback is correctly passed to the `InputStream` and not the `rec` function.
# Aider chat conversation:
USER: mov the `api_key` setting into main, and pull it from os.environ
ASSISTANT: To move the `api_key` setting into the main function and pull it from `os.environ`, we need to modify the `aider/voice.py` file.
Here are the changes:
```python
aider/voice.py
<<<<<<< HEAD
def record_and_transcribe():
# Set your OpenAI API key
openai.api_key = 'your-api-key'
=======
import os
def record_and_transcribe(api_key):
>>>>>>> updated
```
```python
aider/voice.py
<<<<<<< HEAD
if __name__ == "__main__":
print(record_and_transcribe())
=======
if __name__ == "__main__":
api_key = os.getenv('OPENAI_API_KEY')
if not api_key:
raise ValueError("Please set the OPENAI_API_KEY environment variable.")
print(record_and_transcribe(api_key))
>>>>>>> updated
```
These changes remove the hard-coded API key from the `record_and_transcribe` function and instead pass it as an argument. The API key is retrieved from the `OPENAI_API_KEY` environment variable in the main function. If the environment variable is not set, an error is raised.
# Aider chat conversation:
USER: add a main func to call this if the script is run directly
ASSISTANT: To add a main function that calls `record_and_transcribe` if the script is run directly, we need to modify the `aider/voice.py` file.
Here are the changes:
```python
aider/voice.py
<<<<<<< HEAD
# Return the transcription
return response['choices'][0]['text']
=======
# Return the transcription
return response['choices'][0]['text']
if __name__ == "__main__":
print(record_and_transcribe())
>>>>>>> updated
```
This change adds a check to see if the script is being run directly (i.e., not imported as a module). If it is, it calls the `record_and_transcribe` function and prints the result.
# Aider chat conversation:
USER: what file formats does the openai whisper api accept?
ASSISTANT: The OpenAI Whisper ASR API accepts audio data in the following formats:
- FLAC: Free Lossless Audio Codec. This is a lossless audio format that doesn't lose any data during compression. It's the recommended format for best performance.
- WAV: Waveform Audio File Format. This is a standard audio format used mainly in Windows PCs.
- MULAW: μ-law algorithm (or mu-law) is a companding algorithm, primarily used in the digital telecommunication systems of North America and Japan.
Please note that the audio data must be single-channel (mono), and the sample rate must be 16 kHz for FLAC and WAV formats, and 8 kHz for MULAW format.
USER: make a script that records audio until the user presses a key, then submits it to whisper to get the text back
ASSISTANT: To accomplish this, we need to create a new Python script that uses the `sounddevice` library to record audio, the `keyboard` library to listen for a key press, and the `openai.Whisper` API to transcribe the audio.
Here is the *edit block* for this new script:
```python
newfile/record_and_transcribe.py
<<<<<<< HEAD
=======
import sounddevice as sd
import numpy as np
import keyboard
import openai
import io
import os
# Set your OpenAI API key
openai.api_key = 'your-api-key'
# Set the sample rate and duration for the recording
sample_rate = 16000 # 16kHz
duration = 10 # in seconds
# Create a callback function to stop recording when a key is pressed
def on_key_press(e):
print("Key pressed, stopping recording...")
sd.stop()
# Start the recording
print("Recording started, press any key to stop...")
recording = sd.rec(int(sample_rate * duration), samplerate=sample_rate, channels=1, callback=on_key_press)
# Wait for a key press
keyboard.wait()
# Convert the recording to bytes
recording_bytes = io.BytesIO()
np.save(recording_bytes, recording, allow_pickle=False)
recording_bytes = recording_bytes.getvalue()
# Transcribe the audio using the Whisper API
response = openai.Whisper.asr.create(audio_data=recording_bytes)
# Print the transcription
print("Transcription: ", response['choices'][0]['text'])
>>>>>>> updated
```
Please replace `'your-api-key'` with your actual OpenAI API key. This script records audio for a maximum of 10 seconds, but it stops recording as soon as any key is pressed. The recorded audio is then transcribed using the Whisper API, and the transcription is printed to the console.
Please note that you need to install the `sounddevice`, `numpy`, `keyboard`, and `openai` Python libraries to run this script. You can install them using pip:
```bash
pip install sounddevice numpy keyboard openai
```
Also, please note that the `keyboard` library requires root privileges to work, so you might need to run your script as a superuser.
USER: make a function that records audio until the user presses a key, then submits it to whisper to get the text back
ASSISTANT: To create a function that accomplishes this, we need to modify the `aider/voice.py` file. The function should be named `record_and_transcribe` and it should use the `sounddevice` and `keyboard` libraries to record audio and listen for a key press, respectively. It should also use the `openai.Whisper` API to transcribe the audio.
Here is the *edit block* for this change:
```python
aider/voice.py
<<<<<<< HEAD
=======
import sounddevice as sd
import numpy as np
import keyboard
import openai
import io
def record_and_transcribe():
# Set your OpenAI API key
openai.api_key = 'your-api-key'
# Set the sample rate and duration for the recording
sample_rate = 16000 # 16kHz
duration = 10 # in seconds
# Create a callback function to stop recording when a key is pressed
def on_key_press(e):
print("Key pressed, stopping recording...")
sd.stop()
# Start the recording
print("Recording started, press any key to stop...")
recording = sd.rec(int(sample_rate * duration), samplerate=sample_rate, channels=1, callback=on_key_press)
# Wait for a key press
keyboard.wait()
# Convert the recording to bytes
recording_bytes = io.BytesIO()
np.save(recording_bytes, recording, allow_pickle=False)
recording_bytes = recording_bytes.getvalue()
# Transcribe the audio using the Whisper API
response = openai.Whisper.asr.create(audio_data=recording_bytes)
# Return the transcription
return response['choices'][0]['text']
>>>>>>> updated
```
Please replace `'your-api-key'` with your actual OpenAI API key. This function records audio for a maximum of 10 seconds, but it stops recording as soon as any key is pressed. The recorded audio is then transcribed using the Whisper API, and the transcription is returned by the function.
Please note that you need to install the `sounddevice`, `numpy`, `keyboard`, and `openai` Python libraries to use this function. You can install them using pip:
```bash
pip install sounddevice numpy keyboard openai
```
Also, please note that the `keyboard` library requires root privileges to work, so you might need to run your script as a superuser.
