Datasets:
Deprecation Notice — Korean & Hindi Two-Language Pack (v1.0)
Status: Deprecated. Please use the v1.1 replacement files in this folder.
What was deprecated and why
The original two-language release (korean_cv24.parquet and hindi_cv24.parquet, published as
Alexandria Voice Corpus — Korean & Hindi Macro-Prosody Telemetry) was built on a pipeline
with three unresolved quality-gate bugs. Those files should no longer be used.
Bugs present in the v1.0 files
BUG-029 — Missing C50 floor on T2 gate The STUDIO tier (T2) admitted clips regardless of C50 (room clarity) as long as SNR was sufficient. This allowed high-reverb recordings — rooms with significant echo or reflected sound — to be stamped STUDIO. Affected features include spectral tilt, HNR, CPPS, and F0 smoothness, all of which degrade in reverberant conditions. The fix requires C50 ≥ 20 dB in addition to SNR for any clip to reach T2.
BUG-030 — Missing speech-ratio floor on T2 gate The T2 gate had no minimum speech density requirement. Clips where more than 70% of the audio was ambient room tone or silence were admitted to STUDIO tier. This inflated articulation_rate and distorted nPVI values for those clips. The fix requires speech_ratio ≥ 0.30 for any clip to reach T2.
High-SNR C50 bypass A fast path for loud recordings (SNR ≥ 35 dB) bypassed the C50 check entirely. A recording made in a tiled bathroom at high volume could clear the SNR bar but still contain severe reverb. This path was patched to apply the same C50 ≥ 20 dB and speech_ratio ≥ 0.30 requirements as the standard T2 path.
The full T2 gate post-fix requires all three simultaneously: SNR ≥ 25 dB and C50 ≥ 20 dB and speech_ratio ≥ 0.30.
What to use instead
This folder contains corrected replacement files for both languages, plus five additional languages, as the v1.1 multilingual pack:
| File | Language | Rows | Notes |
|---|---|---|---|
korean_cv24.parquet |
Korean | 6,805 | Replaces v1.0 (was 6,998 — difference is T2 gate correction + k5 suppression) |
hindi_cv24.parquet |
Hindi | 18,435 | Replaces v1.0 (was 18,447) |
hebrew_cv24.parquet |
Hebrew | 5,455 | New in v1.1 |
manx_cv24.parquet |
Manx | 6,579 | New in v1.1 |
tzeltal_cv24.parquet |
Tzeltal | 5,585 | New in v1.1 |
mdn_cv24.parquet |
Maguindanao | 5,536 | New in v1.1 |
lasi_cv24.parquet |
Lasi | 10,435 | New in v1.1 |
See README.md in this folder for full dataset documentation.
Row count differences between v1.0 and v1.1
The v1.0 Korean file had 6,998 rows; v1.1 has 6,805 (-193 net). The v1.0 Hindi file had 18,447 rows; v1.1 has 18,435 (-12 net).
The reduction reflects two changes applied together:
- Retroactive tier regrading — clips that passed the old T2 gate but fail the corrected gate (missing C50 or speech_ratio) were downgraded to T3 and excluded from export.
- K-anonymity suppression — the k=5 quasi-identifier check is unchanged, but the regraded tier distribution shifted some group sizes below threshold.
Neither the source audio nor the Mozilla filenames have changed. The underlying clips are the same; what changed is which clips meet the corrected STUDIO quality standard.