arg-spanish-tts

Unified, deduplicated speech corpus for Argentine Spanish (es-AR), built by merging three public datasets and stripping cross-source duplicates. All audio is resampled to 24 kHz mono. Intended for multi-speaker TTS pre-training before fine-tuning to a target voice (a workflow that generally beats single-speaker training from scratch).

Stats

Metric	Value
Rows	10,747
Total audio	12.18 h (43,833 s, 730.5 min)
Unique speakers	65
Sampling rate	24 kHz mono
Gender split	F: 7,419 / M: 3,328

Rows per source (after dedup)

Source	Rows	Notes
giandiego_ar	5,372	Subset of GianDiego/latam-spanish... filtered to nationality == "ar"
ylacombe	5,372	OpenSLR 61 (Google Argentinian Spanish), female + male splits merged
fabricio3g	3	Almost the entire dataset was already present inside ylacombe (see dedup)

Sources

Source key	Upstream dataset	Subset used
giandiego_ar	GianDiego/latam-spanish-speech-orpheus-tts-24khz	nationality == "ar"
ylacombe	ylacombe/google-argentinian-spanish	female + male
fabricio3g	fabricio3g/argentine-tts-4voices	train + eval (4 voices: sofia, valentina, lucas, martin)

Schema

Field	Type	Description
audio	Audio(24000)	Mono, 24 kHz (uniformly resampled with librosa)
text	string	Transcription as provided upstream
speaker_id	string	Original speaker ID from the source dataset
gender	string	f, m or unknown
voice_name	string	Human-readable voice name when available (fabricio3g), otherwise equals speaker_id
source	string	giandiego_ar, ylacombe, or fabricio3g

Deduplication

Cross-source overlap is real, especially because fabricio3g/argentine-tts-4voices is a curated 4-voice subset of the same Google Argentinian Spanish corpus that ylacombe/google-argentinian-spanish ships. Two dedup criteria were applied sequentially while walking the concatenated set in source-priority order (giandiego_ar -> ylacombe -> fabricio3g):

1. Audio hash dedup - SHA-256 over the first 8,000 samples of the decoded float32 array (post-resample to 24 kHz). Rows whose audio fingerprint had already been seen are dropped.
2. (text, speaker_id) dedup - the normalized transcription (lower-cased, stripped of accents and punctuation, whitespace-collapsed) paired with speaker_id. Rows where the same speaker uttered the same normalized text are dropped.

Numbers

Stage	Rows
Concatenated input	12,070
Removed by audio-hash dedup	-589 (almost the full fabricio3g set, confirmed subset of ylacombe)
Removed by (text, speaker_id) dedup	-734
Final	10,747

This is intentionally conservative: a one-shot audio hash collision is very unlikely on 8,000 float32 samples, so a hit is treated as the same recording rather than coincidence.

Suggested use

For TTS fine-tuning that generalises before targeting a single voice:

1. Phase 1 - multi-speaker pre-train with the full set (12 h, 65 speakers).
2. Phase 2 - single-voice fine-tune with your target speaker (10-40 min of clean audio is usually enough on top of a multi-speaker base).

Do not collapse multiple speaker_ids into one embedding during phase 1

• the whole point is to keep speakers separate so the model learns speaker-independent prosody.

License

CC-BY 4.0, inherited from the upstream sources. Refer to each original dataset for full attribution requirements.

Build pipeline

Reproducible scripts that built this dataset are at Kukedlc/tts (build_arg_spanish_tts.py) - single-pass download with HF_HUB_ENABLE_HF_TRANSFER=1, resample with librosa, dedup as described above, push with datasets.push_to_hub.