Live transcription PoC with the Whisper model (using faster-whisper) in a server (restapi) - client (gradio ui/cli) setup where the server can handle multiple clients.

(Server is running separately making it usable with any client side code.)

Sample with a Macbook Pro (M1)

(🔈 sound on, faster-whisper package, base model - latency was around 0.5s)

• $ pip install -r requirements.txt
• $ mkdir models

• Before running the server.py modify the parameters inside the file

# Start the server (RestAPI)
python server.py

# --------------------------------

# Start the Gradio interface on localhost (HTTP)
python ui_client.py

# Start the Gradio interface with their sharing - this way the it'll be HTTPS without the need of certs
SHARE=1 python ui_client.py

# Start the Gradio interface with your own certs
SSL_CERT_PATH=<PATH> SSL_KEY_PATH=<PATH> python ui_client.py

python server.py
python cli_client.py

There are a few parameters in each script that you can modify

This beautiful art will explain this:

- step = 1
- length = 4

$t$ is the current time (1 second of audio to be precise)

------------------------------------------
1st second: [t,   0,   0,   0] --> "Hi"
2nd second: [t-1, t,   0,   0] --> "Hi I am"
3rd second: [t-2, t-1, t,   0] --> "Hi I am the one"
4th second: [t-3, t-2, t-1, t] --> "Hi I am the one and only Gabor"
5th second: [t,   0,   0,   0] --> "How" --> Here we started the process again, and the output is in a new line
6th second: [t-1, t,   0,   0] --> "How are"
etc...
------------------------------------------

• Use a VAD on the client side, and either send the audio for transcription when we detect a longer silence (e.g. 1 sec) or if there is no silence we can fall back to the maximum length.
• Transcribe shorter timeframes to get more instant transcriptions and meanwhile, we can use larger timeframes to "correct" already transcribed parts (async correction)