Transcribe an audio-stream in almost real time using OpenAI-Whisper. Monitor it for specific terms in the transcribed text using fuzzy-matching. Trigger an alarm via Signal messenger when your terms are mentioned.
Run pip install -r requirements.txt to resolve dependencies.
Consists of three parts:
save_stream.pysaves .mp3 files in chunks of 30sec from an audio streamtranscribe.pytranscribes each incoming audio chunk using OpenAI-Whisper. Then, it uses fuzzy matching to monitor the spoken word for specific terms. On match, it callsmsg_group_via_signal.shmsg_group_via_signal.shrelays the alarm message to the signal-cli tool which messages a group on the Signal messenger
Take a look at the files to configure them. For example, you can set durations during which the stream is monitored.
Per default the small OpenAI-Whisper model is used. Thus, the transcription quality is decent but not perfect. Hence, we use fuzzy-matching to monitor for our terms, so that we reduce false-negatives (but increase false-positive) alarms. The benefit of the small model being that it works in almost real-time even on a CPU-only machine with mediocre specs (on AWS a c5a.large EC2 instance is sufficient). There will be some delay because of the 30sec chunks and because inference takes some time, but it's fast enough to process all audio without falling behind. With better specs / GPU you can increase model size for better quality transcriptions or reduce latency.
My very specific use case: My sports team participated in a contest hosted by a local radio station that went on for a couple of weeks. If the name of our team was mentioned on-air, we had a couple of minutes to call the radio station in order to win a prize. This was my alternative to listening to shitty music.