Request sample access
This repository is publicly accessible, but you have to accept the conditions to access its files and content.
This repository contains the dataset specification, annotation schema, and sample metadata files. The production dataset (20,000 hours of 48kHz conversational speech across 30+ languages) is rights-cleared and delivered directly under a commercial license. Approved requesters get the full metadata schema and data dictionary in this repository, and can request a review package with real audio samples, transcripts, and QA summaries. Requests are reviewed within 1 business day.
Log in or Sign Up to review the conditions and access this dataset content.
TTS Voice Dataset
20,000 hours of high-fidelity 48kHz conversational audio across 30+ global, regional, and underrepresented languages, built for text-to-speech, voice cloning, and multilingual speech AI.
This repository contains the full technical specification, annotation schema, and sample metadata files (Parquet). The production dataset is rights-cleared and delivered directly to buyers. Request access to see the full schema and get real audio samples.
Overview
The TTS Voice Dataset is a 20,000-hour collection of high-fidelity 48kHz conversational audio recorded globally across numerous languages, regions, accents, and speaker profiles. It is designed for text-to-speech, voice cloning, speech synthesis, conversational AI, speaker modeling, multilingual speech training, and voice agent development.
The collection emphasizes natural spoken dialogue rather than scripted or robotic recordings, which is what makes it useful for building expressive, human-like voices for real conversational products. The 48kHz audio quality supports high-fidelity voice modeling, while the conversational recordings teach models realistic pacing, tone, turn-taking, emotion, pronunciation variation, and speaker expressiveness.
At a glance
| Total volume | 20,000 hours |
| Sample rate | 48kHz, WAV/PCM where available |
| Bit depth | 16-bit or 24-bit depending on subset |
| Languages | 30+ global, regional, and underrepresented |
| Speech type | Natural conversational dialogue, not only read speech |
| Speakers | Diverse age buckets, genders, accents, native and non-native profiles where available |
Language distribution
| Tier | Share | Hours | Languages |
|---|---|---|---|
| English + major accents | 30% | 6,000 | American, British, Canadian, Australian English |
| Major commercial languages | 35% | 7,000 | Spanish, Mandarin Chinese, Hindi, Arabic, Portuguese, French, German, Japanese, Korean, Italian, Russian |
| High-growth regional | 20% | 4,000 | Indonesian, Vietnamese, Thai, Turkish, Bengali, Urdu, Tamil, Telugu, Marathi, Dutch, Polish, Malay, Swahili |
| Low-resource / long-tail | 15% | 3,000 | Tagalog, Uyghur, Hmong, Cebuano, Tamazight, Tibetan, Sepedi, Guarani, Madurese, Bouyei |
Long-tail coverage is the hard part of multilingual voice sourcing; this dataset was structured to balance commercial usefulness with genuine linguistic diversity.
Technical specifications
- Annotation coverage: language, locale, accent, speaker count, transcript availability, diarization availability, quality score, noise level, topic category
- TTS-specific fields: clean speech seconds, expressive speech seconds, text normalization availability, pronunciation metadata where available
- Audio QA metrics: SNR, loudness LUFS, clipping percentage, silence percentage, background noise level, transcript quality score where available
- Buyer evaluation metrics: transcript WER by subset where available, diarization error rate by subset where available, speaker consistency, pronunciation coverage, phoneme coverage, expressive range
- Delivery format: WAV audio, CSV metadata, optional JSON transcripts and diarization files
Metadata schema
The gated file annotation_schema.json in this repository contains the full per-recording metadata schema with an illustrative example record, including audio QA fields (SNR, LUFS, clipping, silence) and TTS-specific fields (clean/expressive seconds, normalization availability).
The samples/ folder holds sample metadata in Parquet format: index_sample.parquet (item-level) and utterances_sample.parquet (event-level), both conforming to this schema. Values are generated to illustrate structure and field distributions; production records ship in buyer review packages.
How to evaluate this dataset
- Request access using the form above. Requests are reviewed within 1 business day.
- On approval you get the gated files in this repository: full metadata schema, sample metadata Parquet files, data dictionary, and access instructions.
- Request a review package and we deliver real 48kHz audio samples in your target languages, transcripts, metadata CSVs, QA summaries, and licensing documentation within 2 business days.
All samples are delivered with structured CSV metadata and JSON annotation files where available. Buyer review packages include representative media files, metadata samples, annotation schema, QA summaries, and data dictionary documentation.
Licensing
The production dataset is rights-cleared for commercial AI training and licensed directly by Datoric, with documented contributor consent and chain-of-custody. Subset (by language, accent, or hours), exclusive, and custom-collection options are available.
About Datoric
Datoric supplies rights-cleared, spec-exact training data for frontier AI labs and enterprise model teams: expressive multilingual voice, computer-use traces, egocentric and industrial video, human manipulation data, robot episodes, and gameplay trajectories. We also run managed collection pipelines for custom specifications.
Contact: nikhil@arzule.com
- Downloads last month
- -