Softcatala
Mic input for live transcription
After some upgrades and configuration changes the live transcription stopped working. My setup was working okay, but for unknown reason stopped capturing the mic input. Hence upgraded Jetson Orin Nano developer kit 4b to JetPack 5.1.3.
The live input device doesn't seem to capture audio same way as arecord. If I do a test recording so:
arecord -D usbmic test.wav Recording WAVE 'test.wav' : Signed 16 bit Little Endian, Rate 8000 Hz, Mono
the recorded audio is fine. The audio file transcribes correctly.
If I then try live transcription using the same device so:
whisper-ctranslate2 --live_transcribe True --live_input_device 27 ....etc
the process starts okay: Live stream device: usbmic Listening.. (Ctrl+C to Quit)
But that's all. Nothing happens. Seems the capture is working differently from record?
Is live transcribing working for others? If so, please give some setup hints!
Can confirm on Mac as well. Feeding in mp3 works but live stream doesn't, even though the device is detected.
The built-in microphone of the M-MacBooks is known to have problems with the input volume with various programs. Sometimes a sudo killall coreaudiod helps for a while, but not here. In fact, the threshold can be lowered, then it works for me:
whisper-ctranslate2 --live_transcribe True --live_volume_threshold 0.01
This worked for me also on a Linux box. I was getting no output. Named my mic device and it reported it properly, so, that wasn't it. the threshold (for me) at 0.01 picked up crickets farting. I bumped it up to 0.025 and it is better. This is the first faster-whisper setup I have used that actually works. I am getting no outrageous hallucinations with this version. I have used VAD on others, too. I inserted rms threshold and all kinds of checks but still got misread results and crazy nonsense with no input from mic at all.
