calls-train-sonus

ASR training clips from KejueAI/sonus-sip-calls (internal client call recordings, private). 642.4 h kept.

Sonus voice-agent calls (stereo). The AI agent always owns one channel; that whole channel is dropped per call. Clips are cut from the customer channel only, so no crosstalk.

Labels = Soniox stt-async-v5 transcripts (clean verbatim). Clip bounds = Soniox speaker-turn times + 0.15 s pad, trimmed to Silero-VAD speech + 0.25 s. No CTC aligner (avoids the digit-damage bug of the old pipeline). Turns > 25 s were dropped (no word timestamps to split them).

Filters: bot side removed; no detectable speech; speaking rate outside 2-30 chars/s; duration outside 0.5-30 s. Dropped: speaking_rate_h 18.1 h, no_speech_h 20.5 h, bot_h 1389.1 h.

accent column: English speech -> CommonAccent ECAPA vote per (call, speaker) when >= 60 % agree, else english_unknown; a speaker who also speaks Arabic in the same call -> arabic (the accent model has no Arabic class). Arabic speech -> call dialect from manifest_v2 (Gemma text classification). Other scripts -> other_language.

Not a benchmark: these calls overlap the existing KejueAI/asr-* training repos.

Format: parquet, 500-row row groups (Cohere training format). Columns: audio (16 kHz mono FLAC bytes), id, text, duration, accent, speaker_id (call_speaker_channel), gender (empty), source, lang, reference_text (empty), meta (JSON: client, call_id, start/end, bot_score, accent_votes, call_dialect, ...).

client	accent	hours
Jo Academy	levantine	584.4
Jo Academy	other_language	47.0
Jo Academy	saudi	2.2
Maintenance Manager	saudi	2.0
Jo Academy	arabic	1.5
Jo Academy	arabic_unknown	0.7
Jo Academy	mixed	0.6