Every voice, transcribed and timed.
Synchronized audio with transcripts, speaker turns and event labels, recorded by real speakers across regions, dialects and recording conditions.
Available as- 01 · Dataset
- 5.7M+ transcribed, diarized assets
- 02 · Environment
- Transcription evals per language and accent
Licensed, stocked, ready to ship.
Plus custom collection for what doesn't exist yet.
Audio, text and speech
Nine Indian languages plus Indian English
The largest single language block
Followed by Marathi at 7,121 hours
Hindi, Bengali, Marathi, Malayalam, Tamil and more
Continuous annotation and QA workflows
Speech with its context attached.
Audio ships with what a model needs to learn from it: who spoke, what they said, when, and what happened around them.
- Transcription
- Verbatim and clean transcripts with timestamps at segment or word level.
- Speaker diarization
- Who spoke when, across multi-speaker recordings.
- Event labels
- Non-speech events, noise conditions and acoustic scenes.
- Accent & dialect coverage
- Regional speech patterns, emotional tone and real recording conditions.
- Storytelling & podcasts
- Long-form fiction and non-fiction audio for natural prosody.
- Custom collection
- Scripted or spontaneous speech collected to your spec.
Segment-level record. Values are illustrative.
From consented voice to transcript.
One call, through the five stages every recording goes through before it ships.
Record & consent
Real speakers are recruited across regions and dialects and consent before recording. Device and room conditions are logged with every file.
- Who
- Speakers
- Gate
- No consent, no recording kept
Transcribe
Native-speaker transcribers write what was said, verbatim, with timestamps, speaker turns, fillers and code-switching kept.
- Who
- Native transcribers
- Gate
- Second transcriber checks a sample
Redact
Names, numbers and addresses spoken aloud are bleeped in the audio and replaced in the transcript at the same timestamps.
- Who
- Model + reviewer
- Gate
- Every redaction checked by ear
Label & QA
Speaker turns, overlaps and non-speech events are labelled, and a reviewer re-listens to a sample to measure accuracy.
- Who
- Reviewers
- Gate
- Sample accuracy below target is redone
Package
Audio ships with its transcript, speaker file and metadata, in the formats speech pipelines read.
- Who
- Delivery
- Gate
- Metadata per recording
Built for frontier model tasks.
Scope, review criteria and delivery format are agreed per engagement, from the first sample to the final eval.
- Speech recognition
- ASR for languages and accents underserved by public data.
- Text to speech
- Clean, natural long-form speech for voice synthesis.
- Voice agents
- Real conversational audio for call-centre and assistant agents.
- Audio understanding
- Event and scene labels for multimodal models.
- Speech evaluation
- Held-out, expert-checked test sets per language.
What buyers ask us first.
Something else? Ask a specialist →
01Which languages are covered?
Hindi, Bengali, Marathi, Malayalam, Tamil, Gujarati, Kannada, Telugu and Indian English in the licensed catalogue, with custom collection for other languages.
02How is transcription quality checked?
Native-speaker reviewers check transcripts and labels against agreed accuracy targets, with sampling and escalation rules scoped per project.
03How many hours are there per language?
37,847 hours in total: Indian English 7,466, Hindi 7,365, Marathi 7,121 and Malayalam 5,560, with Bengali, Tamil, Gujarati, Kannada and Telugu sharing the remaining 10,000+ hours.
04Is it read speech or real conversation?
Real conversation, from speakers across regions, dialects, devices and recording conditions, including code-switched speech such as Hindi mixed with English. Hesitations, interruptions and overlapping turns are kept, because deployed voice AI has to handle them.
05Can it be used for text to speech?
Yes. Long-form storytelling and podcast audio carries natural prosody for TTS, and custom collection covers scripted speech when you need specific speakers, scripts or recording setups.
06Can you collect custom speech data?
Yes. We scope speakers, demographics, scripts, environments and devices, then agree collection and validation requirements.
Same standard in every domain.
Egocentric video
First-person footage with action and object labels for physical AI.
- 01 Dataset
- 200K+ hrs synced video, IMU, action labels
- 02 Environment
- Action and manipulation benchmarks
Company operations
Workflows, approval trails and legal documents for enterprise agents.
- 01 Dataset
- 52+ companies' ordered approval trails
- 02 Environment
- Agent tasks over approvals, operator-graded
Conversational
Multi-turn conversations, ranked and reviewed for assistant training.
- 01 Dataset
- Multi-turn dialogue, intent-labelled
- 02 Environment
- Multi-turn evals, rubric-graded
Private code
Repositories, commits, pull requests and review history.
- 01 Dataset
- 3B+ lines with commit, PR and review history
- 02 Environment
- Agentic engineering tasks, engineer-graded
Healthcare
Clinical notes, records and imaging, de-identified and expert-reviewed.
- 01 Dataset
- 10M+ de-identified, patient-linked records
- 02 Environment
- Clinical case tasks, clinician-graded
CAD & 3D
Drawings, assemblies and meshes with validated geometry.
- 01 Dataset
- 5M+ typed, native-format CAD files
- 02 Environment
- Part classification and rebuild checks
See real audio & speech data first.
A representative sample under NDA, in the format your pipeline reads. Run it through your own evals first.
- Sample records, as delivered
- Datasheet and schema
- Review criteria
- WAV · FLAC · MP3
{"id":"hi_018422","lang":"hi-IN","dur_s":12.8,
"speakers":2,"sr_hz":16000,
"segments":[{"spk":"S1","t":[0.0,4.6],
"text":"[NAME] ji, order kab tak aayega?"}],
"events":["phone_ring@7.2"],
"pii":"masked","wer_review":"passed"}