Numo Indic Speech is free to use under CC BY 4.0, including commercially. Share a few details to get access. Poseidon may contact you about related datasets. See our Privacy Policy at https://www.psdn.ai/privacy

Numo Indic Speech Dataset - Open source

Numo Indic Speech is a quality-filtered, crowdsourced speech corpus collected by Poseidon AI, Inc. It contains 353,349 recordings totaling 3,845.5 hours across Bengali, Hindi, Tamil, and Telugu. The corpus consists of scripted, single-speaker read speech for automatic speech recognition (ASR), with reference transcripts, speaker metadata, and automated validation metrics for every recording.

Corpus	3,845.5 hours · 353,349 recordings
Languages	Bengali (`bn`), Hindi (`hi`), Tamil (`ta`), and Telugu (`te`)
Speech type	Scripted, single-speaker read speech
Speaker metadata	speaker ID, age, gender, country, and region when available
Topics	Collection prompt or topic label for every recording
Transcripts	Original reference transcript for every recording
Audio	Full-resolution WAV embedded in Parquet · sample rate and channel count reported per recording
Validation	Automated per-recording word error rate (WER), character error rate (CER), and synthetic-speech suspicion score
Integrity	SHA-256 checksum published for every embedded audio object

Language configurations

Configuration	Language	Recordings	Hours
`bengali`	Bengali (`bn`)	57,628	745.1
`hindi`	Hindi (`hi`)	101,450	1,314.6
`tamil`	Tamil (`ta`)	61,783	561.6
`telugu`	Telugu (`te`)	132,488	1,224.2
Total	Four languages	353,349	3,845.5

Each language is published as an independent Hugging Face configuration with one data split. This lets users stream or download only the language they need.

Highlights

Large-scale Indic read speech. More than 3,800 hours of scripted, single-speaker recordings across four languages.
Naturally multilingual transcripts. The corpus retains code-switching and commonly used English terms present in the original reading prompts.
Rich speaker metadata. Recordings include a pseudonymous speaker identifier, with age, gender, country, and region provided when available.
Quality filtered. Automated validation measures transcript alignment with WER and CER and screens recordings for synthetic-speech suspicion.
Ready to use and verify. Full-resolution WAV audio and metadata are embedded together in typed Parquet shards, with a SHA-256 checksum for every audio object.

Quick start

Select a language configuration and stream it without downloading the complete corpus:

from datasets import load_dataset

# Authenticate first with `hf auth login` if access is required.
dataset = load_dataset(
    "psdn-ai/numo-indic-speech",
    "bengali",
    split="data",
    streaming=True,
    token=True,
)

sample = next(iter(dataset))

audio = sample["audio"]
audio_samples = audio.get_all_samples()
audio_samples.data
audio_samples.sample_rate

sample["transcript"]
sample["speaker_country"]
sample["wer"], sample["cer"]
sample["synthetic_suspicion_score"]

Use "hindi", "tamil", or "telugu" in place of "bengali" to load another language. Removing streaming=True materializes that complete language configuration and downloads its embedded full-resolution audio.

Dataset structure

Each row represents one recording and includes embedded audio, its reference transcript, speaker metadata, technical audio properties, automated validation metrics, and an integrity checksum. Audio is stored as a Hugging Face Audio feature and decoded lazily when accessed.

Data fields

All floating-point fields are stored as float64 values rounded to two decimal places.

Field	Type	Description
`audio`	`Audio`	Full-resolution embedded WAV audio, decoded lazily by the Hugging Face `datasets` library.
`audio_id`	`string`	Stable identifier unique to the recording.
`language`	`string`	ISO 639-1 language code: `bn`, `hi`, `ta`, or `te`.
`topic`	`string`	Collection prompt or topic associated with the recording.
`duration`	`float64`	Audio duration in seconds, derived from the WAV file.
`sample_rate_hz`	`int32`	Audio sample rate in hertz.
`num_channels`	`int8`	Number of audio channels.
`audio_format`	`string`	Audio file format. Always `wav`.
`transcript`	`string`	Original reference transcript read by the speaker.
`speaker_id`	`string`	Deterministic pseudonymous speaker identifier, stable across recordings from the same speaker profile.
`speaker_age`	`int32`	Speaker age in years, when available.
`speaker_gender`	`string`	Speaker gender. Allowed values include `female`, `male`, and `prefer-not-to-say`.
`speaker_country`	`string`	Country name or code provided in the source speaker profile.
`speaker_region`	`string`	Country subdivision or region code, when available; otherwise null.
`wer`	`float64`	Word error rate calculated by comparing an automated transcription with the reference transcript.
`cer`	`float64`	Character error rate calculated against the reference transcript.
`synthetic_suspicion_score`	`float64`	Automated synthetic-speech suspicion score from 0 to 1; higher values indicate greater suspicion.
`audio_sha256`	`string`	SHA-256 checksum of the embedded audio bytes.

About Poseidon

Poseidon builds the training data for physical AI. Machines learn physical skill the way apprentices always have, by watching skilled people work, and that footage has to be created on purpose and to specification. Poseidon runs that supply chain: it recruits skilled people and real workplaces around the world, records their work to exact lab specifications, and refines the footage into model-ready datasets.

License and Access

Numo Indic Speech is released under the CC BY 4.0 license. You may use, share, and adapt the dataset, including for commercial purposes, provided that appropriate credit is given, a link to the license is included, and any changes are indicated.

When you use Numo Indic Speech:

Credit Poseidon AI, Inc. (see Citation)
Link to the license: CC BY 4.0
Note any changes you made

Not what you're looking for? Request a custom dataset

Citation

If you use this dataset, please cite:

@misc{poseidonai2026numoindic,
  title        = {{Numo Indic Speech Dataset}},
  author       = {{Poseidon AI, Inc.}},
  year         = {2026},
  version      = {1.0},
  publisher    = {Hugging Face},
  url          = {https://huggingface.co/datasets/psdn-ai/numo-indic-speech},
  howpublished = {\url{https://huggingface.co/datasets/psdn-ai/numo-indic-speech}}
}

Downloads last month: 38

Total file size:

1.23 TB