Licence of test_data, and Central Atlas sources — for a personal Tachelhit learning app
Hello, and thank you for publishing these datasets and the ASR benchmark.
I am building a small, personal, non-commercial app to learn Tachelhit / Central Atlas Tamazight myself (https://tachelhit-drills.vercel.app, source at https://github.com/runningcostabrava/tachelhit-drills). It makes short listening drills from clips, with the transcription and a Catalan translation, and uses speech recognition and forced alignment to line up audio and text.
Three questions:
Licence.
test_data(1,801 short Tachelhit sentences with audio, Tifinagh, Latin and Arabic script) has no licence file. May I use the clips and transcriptions (a) as a test set to measure which recogniser works best on Tachelhit, and (b) possibly as listening drills inside the app, with credit to MTNRA-Amazigh-IA on every item? I will not redistribute the data or sell anything.Models. Your
test_modelsbenchmark (omniASR, MMS, wav2vec2, Gemma) is very useful. Will any fine-tuned model for Tachelhit become public? I would gladly test it on my own clips and share the results back.Central Atlas sources. The variety I am learning is the one spoken between Azilal and Kelaat M'Gouna. If your YouTube collection pipeline (Amazigh-IA-Pipeline) has a curated list of Central Atlas channels or playlists that you can share, or segmented batches that will be published under an open licence, I would be very grateful.
In return I am happy to share what I have found: channels with on-screen Tifinagh text, music videos with real Tachelhit subtitle tracks, and my measurements of several recognisers (Gemini, MMS, wav2vec2-berber, NeMo FastConformer, Omnilingual) on Central Atlas clips.
Thank you for your work on Amazigh language technology.
Josep (hobby project, no commercial use)