Voice AI still underperforms for most of the world's speakers, and the reason is unglamorous: there simply is not enough high-quality conversational audio in most languages. Benchmarks hide this, because benchmarks are built from the same narrow pools of read speech recorded in quiet rooms.
India is the clearest example. Hundreds of millions of people speak Hindi, Bengali, Marathi, Telugu, Tamil and other languages as their first language, and most switch between their language and English within a single sentence. A voice agent that works in a demo with clean Hindi falls apart on a real customer call in Pune, with Marathi, Hindi and English mixed together over a phone line and a pressure cooker going off in the background.
01Why voice AI underperforms here
Three problems compound each other:
- Scarcity. Public speech corpora for most Indian languages are small, and much of what exists is read speech: people reading scripts aloud, which sounds nothing like conversation.
- Code-switching. Real Indian speech moves between languages constantly. Models trained on monolingual data treat every switch as noise.
- Acoustic reality. Phone-quality audio, crowded rooms, traffic and overlapping speakers are the norm in deployment and rare in training data.
The result is a familiar pattern: word error rates that look acceptable on benchmarks and unacceptable on live traffic. Teams then try to close the gap with fine-tuning on whatever audio they can find, which is usually not enough and rarely the right kind.
02What is in the catalogue
Our speech catalogue spans nine languages, Hindi, Bengali, Marathi, Malayalam, Tamil, Gujarati, Kannada, Telugu and Indian English, with 37,847 hours of real conversational audio in total. The largest blocks:
- Indian English: 7,466 hours
- Hindi: 7,365 hours
- Marathi: 7,121 hours
- Malayalam: 5,560 hours
- Bengali, Tamil, Gujarati, Kannada and Telugu: the remaining 10,000+ hours between them
It sits inside a wider collection of more than 5.7 million audio, text and speech assets, including long-form storytelling and podcast audio used for natural prosody in text-to-speech.
03Conversational, not read-aloud
The single most important property of the catalogue is that it is real conversation, collected from real speakers across regions, dialects and recording conditions. Not read-aloud scripts in a studio.
Conversational audio contains the things models need to handle and read speech does not: hesitations, restarts, interruptions, overlapping turns, regional pronunciation, and code-mixed sentences like asking when an order will arrive in a mix of Hindi and English. It also covers a range of devices and environments, so the acoustic conditions resemble deployment.
For ASR and voice-agent teams targeting the Indian market, this is the difference between a demo and a deployment.
04How the audio is prepared
Audio becomes training data when it 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. Verbatim transcripts preserve fillers and false starts for ASR; clean transcripts suit downstream language tasks.
Diarization and events
Speaker turns across multi-speaker recordings, plus labels for non-speech events, noise conditions and acoustic scenes. A phone ringing seven seconds into a call is information, not just noise.
Native-speaker review
Native-speaker reviewers check transcripts and labels against agreed accuracy targets, with sampling and escalation rules scoped per project. Personal information such as names is masked before delivery.
Data is delivered in the formats speech pipelines expect: WAV, FLAC or MP3 audio, JSON or JSONL transcripts, RTTM for diarization and TextGrid for alignment.
05Measure it per language and accent
An aggregate word error rate across languages hides the problem it is meant to measure. A model can improve on Hindi and Indian English, which dominate most mixed test sets, while getting worse on Kannada or on rural Marathi accents.
We build held-out, expert-checked test sets per language, and per accent where the project calls for it, so teams can see exactly where a model is weak before their users do. For voice agents, we extend this to full conversational evaluations that score task success, not just transcription accuracy.
06Beyond the catalogue
The licensed catalogue covers the nine languages above. For other languages, dialects or domains, such as a specific industry's vocabulary or a particular call-centre workflow, we run custom collection: scoping speakers, demographics, scripts, environments and devices, then agreeing collection and validation requirements before recording starts.
Our audio and speech vertical has the full breakdown, and samples are available per language under NDA.
