Synthetic Indian call-centre agent speech, one row per clip (24 kHz mono FLAC), additive to the earlier asset-probe datasets. Same clip schema, plus native_language, batch, render_variant, part, style, code_mix, script, cut_from, cut_start_s, cut_end_s, clip_kind, prompt_kind, hard_text, asr_cer, timbre_cosine.

96,571 clips, 128.71 h of audio, 12 push(es). Reports: reports/.

Held-out evaluation voices (batch xlang02-heldout, 12 new voices) live only under heldout/ (config xlang02-heldout, split test); the default and per-language configs never include them.

Batch	Clips	Hours
v3-long	3,965	22.54
v3-newvoices	32,189	45.43
v3-pilot	975	1.51
v3-prompts	2,710	6.97
v3-short	52,368	46.03
xlang02-heldout	4,364	6.22

Training data (data/), by part:

Part	Clips	Hours
A	67,626	57.60
B	3,989	22.69
C	4,945	12.51
D	15,647	29.69

Training data (data/), by language:

Language	Clips	Hours
bn-IN	9,110	11.35
en-IN	10,412	15.66
gu-IN	8,973	10.78
hi-IN	11,983	18.52
kn-IN	8,563	11.41
ml-IN	8,860	11.45
mr-IN	8,923	11.01
pa-IN	8,132	9.88
ta-IN	9,136	11.70
te-IN	8,115	10.71

Parts: A short agent replies (clip_kind short_cut, cut from the short_full render named in cut_from at the pause between the reply and the next sentence; one TTS call per full render), B long agent turns (25-60 s, one TTS call), C Hindi / Indian English / Hinglish prompt clips (prompt_kind hi, en, hinglish, hinglish_latin), D new voices (native_language = none). text is exactly the text sent to the TTS. audio_sha256 is the sha256 of the 16-bit PCM WAV for v3-pilot rows and of the FLAC bytes for later batches.

Downloads last month: 8,969

Total file size:

10.7 GB