Note · 27 Jul, 05:14 HKT
第一份交到我哋手上嘅真實錄音係 GSM 6.10，冇一個瀏覽器解得到。
Evaluation audio is scored as 16 kHz mono PCM, and that is not what comes out of a contact centre. The first real recording we were given was a WAV carrying GSM 6.10 — a compressed telephony codec that Chrome refuses outright.
Uploads now convert in two places. The browser handles everything it can already play, and shows the waveform and peak level of the converted audio so an organiser who cannot listen at their desk can still see a silent take or a clipped one. Anything the browser cannot open is converted on the server: A-law and µ-law are decoded directly, and the rest goes to ffmpeg.
Clips can also be corrected now. A transcript fix writes a new version rather than overwriting the old one, because a score computed last week was computed against the version that was current then.
Cantonese Voice Benchmark · Published 27 Jul, 05:14 HKT.