Access to this dataset is reviewed manually. Please describe who you are and what you intend to use the audio for. Requests without a clear, specific purpose will not be approved.

gdrive src05 + src06 — PESQ > 2.4 filtered subset

395 Yoruba speech clips selected from two internal audio-chunk corpora by BiCodec reconstruction quality, shipped with transcripts and with the codec reconstruction of each clip alongside the original.

Yield

Source	Scored	Kept (PESQ > 2.4)	Rate
`gdrive_src05`	1000	158	15.80%
`gdrive_src06`	997	237	23.77%
merged	1997	395	19.78%

Languages — read this before filtering

The two sources are different languages:

Source	Language	Clips
`gdrive_src05`	Hausa	157 (+1 English)
`gdrive_src06`	Yoruba	234 (+3 other)

Per-row language and language_conf columns are included, assigned by fastText lid218e (facebook/fasttext-language-identification). A handful of rows land on implausible labels (English, Irish) — those are short interjections or code-switched fragments where language ID is unreliable. Filter on language_conf if that matters to you.

If you are training a Yoruba model, filter to language == 'yoruba' or source == 'gdrive_src06'. Taking the whole set mixes ~40% Hausa into it.

The wider gdrive family is multilingual

The six source corpora this subset came from are not one language:

Source	Dominant	Share
`source-01`	Nigerian Pidgin	70% (+23% English)
`source-02`	Yoruba	93%
`source-03`	Igbo	99%
`source-04`	Hausa	98%
`source-05`	Hausa	88%
`source-06`	Yoruba	75% (+8% English)

Columns

Audio

audio — the original clip, unmodified, at native 48 kHz
audio_bicodec — the same clip encoded and decoded through BiCodec (Spark-TTS 0.5B; 32 global + 50 semantic tokens/s), 16 kHz. Useful as a concrete upper bound on what any BiCodec-based TTS can reproduce.

Text

text — transcript, joined from the source repo's train/metadata.jsonl by chunk_id. All 395 rows matched; none are missing.
speaker_local, recording_id, source_audio_filename, start_time_seconds, end_time_seconds

Metrics — read the suffix carefully

The _scored15 columns come from a benchmark that truncates every clip to its first 15 seconds before scoring. pesq_scored15 is the column the > 2.4 selection was made on, so it is kept for auditability even though it describes only part of the longer clips.

The _full columns re-score the whole clip and are the ones to trust when judging the audio you actually have.

pesq_scored15, pesq_internal_scored15, stoi_scored15, estoi_scored15, mcd_scored15, spk_cos_scored15, lsd_lo_scored15, lsd_4k8k_scored15, hb_energy_ratio_db_scored15, dur_scored15, n_tokens_scored15
pesq_full, stoi_full, estoi_full, dur_full, n_tokens_full

How the 15 s cap affects the selection

219 of 395 clips are longer than 15 s, so for those the selection score covered only part of the audio. Re-scoring the whole clip:

mean PESQ 2.633 (_scored15) → 2.612 (_full); median change −0.001
355 / 395 (89.9%) still clear 2.4 on the full clip
among the 219 long clips, 82.7% still clear 2.4; their mean change is −0.036

So the cap inflated scores slightly and the effect is concentrated in long clips, but it did not manufacture the selection. If you want a strictly clean cut, filter on pesq_full > 2.4 and you keep 355 clips.

Caveats

mcd_scored15 is on a non-standard scale (natural-log mel, without the ×10/ln10 factor). Comparable within this dataset only — do not compare it against published MCD figures.
A high PESQ means the codec reconstructs the clip well. It is not evidence of a clean recording, a correct transcript, or a single speaker.
spk_cos_scored15 varies widely among kept clips. If speaker consistency matters, filter on it separately — PESQ does not imply it.
Durations run from ~2.7 s to ~37 s. Pipelines that assume bounded segment lengths may need to re-chunk, and re-chunking would require re-aligning text.

Downloads last month: 11

Total file size:

961 MB