Reubencf/goan-konkani-speech

Goan Konkani Speech (Romi)

47,365 audio clips, 108.5 hours of Goan Konkani (ISO 639-3 gom) speech from Goan television news, transcribed in Romi Konkani - Konkani written in the Roman script.

Konkani is a low-resource language with very little public speech data. This is assembled from broadcast news, so it is real spoken Konkani: studio anchors, field reporters, phone interviews, and the Konkani-English code-switching that Goan speakers actually use.

Contents

Clips	47,365
Hours	108.5
Sample rate	24 kHz mono FLAC
Clip length	3-15 s (median ~8 s)
Script	Romi (Roman)

Tiers

Every clip carries a tier field naming the strictest tier it qualified for, so the two values are disjoint, not nested: tier == "stt" is 29,205 clips, NOT the whole corpus.

• For ASR, use everything - all 47,365 clips, 108.5 h. Every row is intelligible speech, and the noisy field audio and code-switching are an asset for ASR rather than a defect. Do not filter on tier here; filtering to "stt" would throw away the 18,160 cleanest clips in the corpus.
• For TTS, filter tier == "tts" - 18,160 clips, 41.6 h of clean single-speaker studio speech, high confidence, boundaries confirmed against detected silence.

The other 29,205 clips (66.9 h) are good speech that did not clear the stricter TTS bar - too noisy, more than one voice, or a boundary that no silence backed up.

Channels

Channel	Clips
Prudent Media Goa	17,412
DD News Goa	12,205
Prime Tv Goa	9,103
Goa 365 TV	8,645

Speaker gender (tts tier)

male	20.9 h
female	20.7 h

Language mix

Tag	Clips
kok-en	27,641
kok	19,724

kok is Konkani; kok-en is Konkani with English words embedded in Konkani grammar (panel-acher, England-antlea), which is normal Goan speech rather than an error.

Splits

Split	Clips
train	42,352
validation	2,977
test	2,036

Splits are assigned by source broadcast, never per clip, so no recording appears in two splits. Clips from one broadcast share a speaker and an acoustic channel, and a clip-level split would leak near-duplicates into the test set.

Usage

from datasets import load_dataset

# AudioFolder serves the whole corpus as one split named "train"; the real
# train/validation/test assignment is the `split` COLUMN, not the loader split.
ds = load_dataset("Reubencf/goan-konkani-speech", split="train")

train = ds.filter(lambda r: r["split"] == "train")
test = ds.filter(lambda r: r["split"] == "test")
tts = ds.filter(lambda r: r["tier"] == "tts")

How it was built

1. Audio downloaded from the source broadcasts and cut into ~100-second windows at detected silences.
2. Each window annotated by Gemini 3.8 Flash via the Batch API, returning sentence-level segments with start/end times, a Romi transcript, and tags for language, audio type, speaker count and gender.
3. Timestamps corrected locally: Gemini proposes boundaries, then each is snapped to real silence found with ffmpeg silencedetect, tracking a rolling drift estimate. Segments whose boundaries stay far from any silence are rejected.
4. Clips cut from the original audio, filtered, and split by broadcast.

Window length matters: over ~300-second windows the model's timestamps drift by tens of seconds while still looking plausible, so windows are kept near 100 seconds and anything above 150 is not trusted for timing.

Known limitations

• Clip boundaries were corrected in this build. An earlier version pushed some clips whose start had been moved later on an extrapolated drift estimate that no detected silence confirmed, which cut the first word off 18.6% of clips. Where a boundary now has no silence backing it, the clip is widened by the size of that unverified correction instead. If you pulled this dataset before, re-pull it.

• The alignment figures below predate that fix and were measured on the older, truncated clips, so treat them as a floor rather than a current reading - a clip missing its first word scored worse partly for that reason. A full re-measurement across every clip is in progress.

• Alignment is good but not perfect. 1,088 clips were checked by re-transcribing the cut audio on its own and comparing against the stored text, after normalising for spelling (Romi has no standard orthography, so zata/jata had to count as one word). Content-word agreement:

  agreement	share of clips
  >= 0.90	10%
  >= 0.75	36%
  >= 0.60	60%
  >= 0.45	78%
  >= 0.30	90%

  The bottom ~10% look genuinely mismatched. The rest of the disagreement is two passes describing the same audio differently, which is expected for an unstandardised script. Treat it as label noise: fine for ASR, worth filtering for TTS - the tts tier scored better than the corpus as a whole.

  Note that snap_residual does not predict alignment quality - agreement was flat across every residual band - so do not filter on it.

• Romi has no standardised orthography. The same word may be spelled differently across clips. Normalising before training is worthwhile.

• Speaker identity is not labelled. video_id plus speaker_gender is the closest proxy; the same anchor recurs across many broadcasts without being linked.

• Not a balanced corpus. It is news speech: formal register, current affairs vocabulary, and one channel contributes more than the others.

• Transcripts are machine-generated and have not been reviewed by a Konkani speaker.

Licensing

The recordings originate from television news broadcasts and the underlying rights belong to the broadcasters. This dataset is shared for research on a low-resource language. Check the terms that apply to you before redistributing or using it commercially.

Fields

Field	Meaning
audio	24 kHz mono FLAC
text	Romi Konkani transcript
duration	clip length in seconds
tier	stt or tts
split	train / validation / test
video_id, channel, source_url	provenance
start, end	position in the source broadcast
language	kok, kok-en, ...
speaker_gender, audio_type, confidence	annotation tags
drift_at_segment, snap_residual	timing-correction diagnostics