Datasetpublic

groxaxo/google-latam-spanish-uniform-vad

Google LATAM Spanish — Uniform VAD and 0.5 s Edges This public derivative contains 15,016 Latin American Spanish utterances from the Google crowdsourced TTS datasets packaged by ylacombe. Female and male audio were freshly exported from the same pinned upstream revisions and passed through exactly the same processing pipeline. Configurations Configuration Train Validation Total Hours including edge padding argentina-female 3,542 379 3,921 4.181… See the full description on the dataset page: https://huggingface.co/datasets/groxaxo/google-latam-spanish-uniform-vad.

sourceHugging Facecc-by-sa-4.0updated 6d agoView on Hugging Face

0likes104downloads

Dataset Card

Google LATAM Spanish — Uniform VAD and 0.5 s Edges

This public derivative contains 15,016 Latin American Spanish utterances from the Google crowdsourced TTS datasets packaged by ylacombe. Female and male audio were freshly exported from the same pinned upstream revisions and passed through exactly the same processing pipeline.

Configurations

Configuration	Train	Validation	Total	Hours including edge padding
`argentina-female`	3,542	379	3,921	4.181
`argentina-male`	1,678	140	1,818	1.956
`chile-female`	1,599	139	1,738	1.948
`chile-male`	2,494	142	2,636	2.786
`colombia-female`	2,220	149	2,369	2.603
`colombia-male`	2,384	150	2,534	2.670
Total	13,917	1,099	15,016	16.145

Each configuration is independently speaker-disjoint between train and validation. Country and gender are never mixed within a configuration.

Uniform processing

Every source clip received the same treatment, without country- or gender-specific branches:

1.Export the original upstream audio as 48 kHz, mono, 16-bit PCM.
2.Detect the first and last speech regions with Silero VAD pinned to commit 60b7ffa243625ebdc1070275a29f18c87843786a.
3.Use probability threshold 0.5, minimum speech 100 ms, minimum silence 100 ms, and 50 ms protective speech context.
4.Retain everything from the first detected region through the last detected region, including internal pauses.
5.Remove only contiguous literal-zero samples at the retained edges.
6.Add exactly 0.500 seconds (24,000 frames) of digital zero at both the start and end.

No fixed dBFS amplitude threshold was used. All 15,016 clips produced a VAD region and passed a full post-processing scan for format, exact edge padding, and manifest duration consistency.

Sources

Country	Dataset	Pinned revision
Argentina	`ylacombe/google-argentinian-spanish`	`a7e9786e3930b57ae95d38859106e4b50444f943`
Chile	`ylacombe/google-chilean-spanish`	`082619e99a37715ab8db45b0bc8188ba98a80130`
Colombia	`ylacombe/google-colombian-spanish`	`18c187b1bfa6f5ada649e4886d02744ea9acafa2`

The original recordings, transcripts, and speaker labels come from those datasets. Thanks to ylacombe for publishing and maintaining them.

Fields

—audio: embedded WAV audio
—id: stable country/gender/speaker/source-index identifier
—text: exact upstream transcript
—speaker_id and speaker_key: speaker identifiers
—country and gender: fixed values for the selected configuration
—source_dataset, source_revision, and source_index: provenance
—duration: processed duration including both 0.5-second edges
—sample_rate: 48,000 Hz
—audio_sha256: checksum of the processed WAV
—processing: pipeline identifier

License

CC BY-SA 4.0, matching the upstream dataset declarations. Preserve upstream attribution when redistributing or adapting this dataset.