๐๐ฅ I honestly didnโt expect this to work this well. I just ran a 600M-parameter Amharic speech-to-text mode entirely on my own laptop, without any cloud service, API key, or per-minute fees. I downloaded the model weights, started a local web server, uploaded an audio file, and got the Amharic transcription directly on the page.
I tested it with a full Amharic news clip and got around 4.5% CER and 18% WER on clean, read speech, which is roughly in line with the performance published by the developer.
The part I found most interesting was the file test. I uploaded a real MP3 file, then took an MP4 video, changed its extension from .mp4 to .mp3 without actually converting it, and uploaded that too. It still produced a similarly clean transcription.
Itโs pretty cool to have a 600M-parameter Amharic speech model running completely locally on my laptop without depending on an external transcription service. ๐ป๐ฅ
๐คจ Don't look the UI.
#AfroDev #HoHe
I tested it with a full Amharic news clip and got around 4.5% CER and 18% WER on clean, read speech, which is roughly in line with the performance published by the developer.
The part I found most interesting was the file test. I uploaded a real MP3 file, then took an MP4 video, changed its extension from .mp4 to .mp3 without actually converting it, and uploaded that too. It still produced a similarly clean transcription.
Itโs pretty cool to have a 600M-parameter Amharic speech model running completely locally on my laptop without depending on an external transcription service. ๐ป๐ฅ
๐คจ Don't look the UI.
#AfroDev #HoHe