This dataset contains machine-transcribed Persian call-center and YouTube audio, including real customer speech (names, phone numbers, order details). Access is granted manually for research on speech recognition only. You must not redistribute the data, attempt to identify or contact any person in it, or use it for any other purpose.

Audio-Encoder-Training-Data

Stage-1 data for adapting the audio encoder of Gemma 4 E2B to Persian ASR (audio -> Soniox transcript, no draft in the prompt). Three configs, one schema. Sensitive: the call-center part contains real customer calls (names, phone numbers, order details). Labels are machine-generated (Soniox stt-async-v5), not human transcripts.

config	rows	hours	train h	val h
movies	65,917	201.4	197.2	4.2
youtube	41,149	199.7	195.7	4.0
callcenter	64,895	260.5	256.6	3.9
total		661.7

movies: Persian series / movie-related YouTube videos (736 videos, 10 channels). Short clips of the same video are stitched into 2-30 s chunks (random gap fill: room tone, edge noise, silence or crossfade). Augmentation: speed, telephone band-limit (70%), codecs (mu-law/A-law/GSM/AMR/Opus), background babble/music, dropouts (word-guarded), gain.
youtube: Persian YouTube channels (2,459 videos, ~97 channels, 9 domains: podcasts, tech reviews, cooking, ...). Tech reviews are kept whole; every other domain is sampled proportionally. Augmentation: speed, pitch, reverb, background babble, line noise, mic EQ, AGC, telephone band-limit (wideband audio only), codecs, short dropouts, gain; at most 3 per chunk (2 for noisy chunks).
callcenter: Persian call-center audio (agent c0 / customer c1, 8 kHz telephony upsampled to 16 kHz): the 150h Soniox-labelled chunks plus test-1k channels cut into 2-30 s chunks (VAD + CTC forced alignment as a quality filter). Each training chunk appears twice with different augmentation (copy_idx 0/1). Augmentation: speed, pitch, reverb, background babble, line noise, mic EQ, AGC, gain (no band-limit, codecs or dropouts: the audio is already telephony); at most 3 per chunk.

Columns (identical in all configs)

chunk_id, video_id, source, copy_idx, clip_ids, clip_pools, clip_starts, clip_ends, gaps, fill_modes, audio (augmented FLAC, 16 kHz mono), duration (of the final audio, seconds), clip_texts, text (Soniox, punctuated), text_norm (training target), clip_confs, conf_min, conf_mean, n_clips, aug_params (JSON of what was applied), split (train / val; the Hub split validation is the val-* files).

text_norm: Arabic letters -> Persian, punctuation removed, ZWNJ kept, digits as ASCII numerals (spoken numbers stay words). Score with a number-folding WER/CER.

Field meaning differs a little per source:

video_id: video id (movies, youtube) or call id (callcenter); clip_pools: QC pool letter / domain / channel (c0 agent, c1 customer).
n_clips: 1-7 stitched clips in movies, always 1 elsewhere; gaps / fill_modes are empty outside movies.
conf_min: worst clip in a stitch (movies); equals conf_mean elsewhere.
callcenter: both copies of a chunk share chunk_id; the key is (chunk_id, copy_idx). Validation rows are clean, one copy.
Validation rows of all configs are not augmented.

Notes

Skip rows with duration > 30 (3 movie rows slightly over the model's 30 s limit).
With datasets >= 4, decoding audio needs torchcodec; or use Audio(decode=False) and decode the bytes with soundfile.
callcenter/heldout_calls.json: test-1k calls kept out of training as full-channel evaluation audio.

Downloads last month: 15

Total file size:

34.9 GB