Server side

• load the model to GPU
• listen to port 8001 for incoming audio file request
  • endpoint = "http://127.0.0.1:8001/transcribe"
• transcribe and translate the audio into Englisch text.

Server side installation

# install pytortch with cuda 11.8
pip3 install 

# install requirements
pip3 install scipy "numpy<2" fastapi[standard] transformers

# other requirements through sudo apt-get install
# TBD, just install anything that is missing

Then, download the model from the Huggingface model hub and save it somewhere here.

• Just save it in this BBAudioTranscription folder
• Be sure to download the whisper-medium model

Server side running

The server side can be deployed in a local machine or in WS.

With a local PC

# just run the python script audio_api.py
python3 audio_api.py

With WS

# In WS, for whisper-medium, fast-api listens port 8001, start a worker
rlaunch -L 8001 --cpu=1 --gpu=1 --memory=8192 -- python3 audio_api.py    

# In local PC, start SSH forwarding
ssh -NL 8001:localhost:8001 <WS-username>@<WS-IP>

Client side

• start recording with "v"
• stop recording with "v" again
• send the audio file to the server
• wait for the response, which is the string of the transcribed text

Client side installation

pip3 install sounddevice pynput pydub scipy "numpy<2" 

Client side running

python3 audio_client.py

---

BB

• HTTP 422 UNPROCESSABLE CONTENT

  • key error (fastapi app arg)

• Noise in the recording

  • High-pass filter with a cut-off frequency of 40Hz

whisper

• sampling rate need to be 16000