Amharic Voice Command AI

I’m building an Amharic speech-recognition system that reliably understands short voice commands on both mobile devices and desktop computers. To reach production quality I need help gathering or generating a robust command-focused dataset, training a low-latency model, and packaging everything so it can run efficiently on-device.

Here’s what I need from you:
• Collect or record clean Amharic voice-command audio with good speaker and acoustic diversity, then label each clip in UTF-8 Amharic text.
• Pre-process the corpus (16 kHz, mono, noise-reduced) and create train/validation/test splits.
• Train and fine-tune an acoustic and language model—Kaldi, Wav2Vec 2.0, Whisper, or similar—so it recognises common control phrases instantly.
• Optimise and export an inference-ready model that runs offline on an average Android handset as well as standard PC CPUs.
• Provide all training scripts, configuration files, and usage instructions in a Git repository so I can reproduce and extend the work.

Acceptance criteria
1. ≤ 10 % word-error rate on a held-out test set of 1–5 second commands.
2. Startup latency under 500 ms and per-command recognition under 150 ms on target hardware.
3. Complete, well-commented source code, data splits, and README delivered.

If you have prior experience with Amharic or other low-resource languages and can meet the above metrics, I’d love to collaborate.