01
Conversation
Studio Conversations
Paired, high-fidelity dialogue captured in controlled acoustic environments.
• Dialogue AI• Voice cloning• TTS
21+ languageschannel-separated · 48kHz WAV
01Voice datasetsavailable for·Model Training·Evaluation
annotate/captureverified
₹94,000 का AC-fridge एक बार में भरें या EMI में convert करवाएँ
Hindi · Banking & Finance14:47 · 16kHz · dual-channel
ma'am मैंने अभी ना recently एक अ, अ fridge buy किया था। AC fridge buy किया था।
अच्छा देखिए अगर है ना अ आप loan करवाना चाहते हैं ना card पर,
Open full conversation166 lines
A
0:02 — 0:02
Hello.
B
0:04 — 0:05
Hello।
A
0:05 — 0:06
हाँ जी ma'am.
B
0:07 — 0:08
हाँ जी।
B
0:09 — 0:11
जी प्रशांत जी।
A
0:11 — 0:12
जी, जी, जी ma'am.
A
0:15 — 0:16
Ma'am अभी--
B
0:15 — 0:16
जी बताइए।
A
0:17 — 0:22
ma'am मैंने अभी ना recently एक अ, अ fridge buy किया था। AC fridge buy किया था।
B
0:24 — 0:24
हाँ।
A
0:25 — 0:32
वो मतलब उसमें जो offer था ना वो था छह महीने पर चौदह percent ma'am था और नौ महीने पर पंद्रह percent था।
A
0:33 — 0:42
तो ma'am मतलब ऊपर से उसमें अ, four double nine थी और उसमें था कि एक बार में आपको चौरानवे हज़ार मतलब निकालना है।
B
0:34 — 0:35
अच्छा।
A
0:44 — 0:53
तो ma'am मुझे ऐसा लग रहा है कि एक बार में जो चौरानवे हज़ार मुझे निकालना है वो थोड़ा सा जैसे एक वो है ना भारी लग रहा है। मतलब हम कह सकते हैं कि ज्यादा amount लग रहा है।
B
0:44 — 0:44
हमम।
B
0:55 — 1:01
अच्छा देखिए अगर है ना अ आप loan करवाना चाहते हैं ना card पर,
A
1:00 — 1:02
जी, जी। जी।
B
1:03 — 1:34
तो आपको अगर loan कराना चाहते हैं ना sir जी, तो छह महीने पर आपका full interest तकरीबन अ अड़तीस सौ के आसपास रहेगा और अगर आप नौ महीने पर कराओगे ना कि नौ महीने की आपको loan करवाने हैं तो आपको छह हज़ार के अ छह हज़ार अ interest लगेगा और आपकी fees जो पाँच सौ है मतलब चार सौ निन्यानवे रुपये है वो अलग से रहेगी। और मतलब अ एक बार वो है ना कि आप एक reversal अ तभी एक बार करने पर आपके कहीं से उधार ले।
A
1:16 — 1:17
जी।
A
1:25 — 1:25
जी।
telephony · 16kHz · dual-channel · real-world noiseBanking & Finance
02——The Problemwhy this matters
India speaks 22+ languages. Quality AI training data for most of them doesn't exist yet.
01CRITICAL
Low resource languages
Many Indic languages lack the data volume needed to train reliable, high-quality models.
02HIGH
Code-mixed speech
Hindi-English, Tamil-English: colloquial patterns are underrepresented and difficult to source at scale.
03HIGH
Domain-specific gaps
Finance, education, and support verticals need specialised conversational data that generic datasets don't cover.
3 gaps identified. Data infrastructure needed.
03——What's differentby design
Real phone calls, briefed for the domain your product actually ships in.
01
Real phone calls
Captured over the PSTN, not studio reads or app uploads. Each speaker is recorded on their own leg.
3files per call
02
Any domain you need
Calls are briefed from a scenario bank we author per vertical, so the data matches your product instead of a catalogue.
900scenarios
03
Code-mixed by default
Nobody was given a script, so languages switch mid-sentence — the point where recognition actually breaks.
84%of utterances
04
Labelled per speaker
Age, gender, city and home region on every voice, so you can report accuracy by subgroup instead of in aggregate.
99cities
Measured on the live corpus. New domains authored on request.
04——Who we serve
The industry changes. The data gap stays the same.
Pick yours. We'll show you the kind of data and evaluation work we do for companies just like you. Real problems, real datasets we've built.
Healthcare
“Patient portals and IVRs that no one understands in their own language.”
- Multilingual IVR trainingNatural speech data for appointment booking, symptom triage, and prescription refills in 10+ Indic languages.
- Clinical NLP datasetsDoctor-patient conversations annotated for entity extraction, intent, and medical terminology.
- Voice bot evaluationNative speakers stress-test healthcare bots for comprehension, safety, and empathy in regional languages.
05 The collection
Five datasets · one standardFrom close-mic clarity to the beautiful mess of everyday speech.
Every collection is consented, documented, and engineered for the trade-offs that appear after deployment — variability, accent breadth, noise, and conversational context.
Conversational datasets
02
Conversation
Real World Conversations
Natural phone and field conversations with the texture, noise, and spontaneity of daily life.
• ASR• Accent coverage• Noise robustness
21+ languagesdual-channel · 16kHz WAV
03
Conversation
Group Conversations
Unscripted multi-party speech with interruptions, overlap, and authentic turn-taking.
• Diarization• Separation• Meeting AI
21+ languagesmulti-channel · 48kHz WAV
Single speaker datasets
04
Single speaker
Short Form Snippets
Purpose-built, studio-grade utterances designed for broad phonetic and expressive coverage.
• TTS• Wake word• Assistants
21+ languagesmono · 48kHz WAV
05
Single speaker
Long Form Narration
Continuous storytelling and explanatory narration for natural, sustained speech generation.
• Audiobooks• Narration• Expressive TTS
21+ languagesmono · 48kHz WAV
06How it workscollect → verify → deliver
From phone call to
training dataset
No apps to download. No complex setup. Just real phone conversations, verified and structured for your AI models.
014:32
Collect
Real phone conversations in native languages across your target domains. No apps needed, just regular phone calls that capture authentic speech patterns.
02
Verify
Every recording passes automated quality checks and human review. Annotation, labeling, and demographic metadata added by trained linguists.
Audio
Language
Quality
03
Deliver
Structured datasets with transcripts, timestamps, speaker labels, and demographics. Ready for your training pipeline.
dataset.csv
1speaker_a, 0:12, "नमस्ते...", hindi, female
2speaker_b, 0:18, "हाँ बोलिए...", hindi, male
07——Evaluationnative speaker testing
Native speakers test what benchmarks can't measure.
Real-world and adversarial scenarios. Failure modes identified before they reach production, not after.
0evaluation types
01
Native speaker validation
Native speakers evaluate model outputs for fluency, naturalness, and accuracy. Judgments that require genuine comprehension, not pattern matching.
02
Adversarial testing
Code-switched queries, jailbreak and safety red-teaming, bias and cultural sensitivity audits across caste, gender, religion, and region.
03
Benchmarking
Compare your model against competitors on language, domain knowledge, and reasoning. See where targeted data investment closes the gap.
04
Analysis reports
Understand the scenarios where your model fails and get actionable suggestions to fix it from our team of domain experts.
08——Languages21 supported
Every major Indian language.
We support every major Indian language, from Hindi to Sanskrit, with dialect-level granularity across all 21.
0languages
हिन्दीHindi
తెలుగుTelugu
മലയാളംMalayalam
தமிழ்Tamil
ಕನ್ನಡKannada
বাংলাBengali
मराठीMarathi
ગુજરાતીGujarati
ਪੰਜਾਬੀPunjabi
ଓଡ଼ିଆOdia
اردوUrdu
हिन्दीHindi
తెలుగుTelugu
മലയാളംMalayalam
தமிழ்Tamil
ಕನ್ನಡKannada
বাংলাBengali
मराठीMarathi
ગુજરાતીGujarati
ਪੰਜਾਬੀPunjabi
ଓଡ଼ିଆOdia
اردوUrdu
हिन्दीHindi
తెలుగుTelugu
മലയാളംMalayalam
தமிழ்Tamil
ಕನ್ನಡKannada
বাংলাBengali
मराठीMarathi
ગુજરાતીGujarati
ਪੰਜਾਬੀPunjabi
ଓଡ଼ିଆOdia
اردوUrdu
हिन्दीHindi
తెలుగుTelugu
മലയാളംMalayalam
தமிழ்Tamil
ಕನ್ನಡKannada
বাংলাBengali
मराठीMarathi
ગુજરાતીGujarati
ਪੰਜਾਬੀPunjabi
ଓଡ଼ିଆOdia
اردوUrdu
অসমীয়াAssamese
मैथिलीMaithili
کٲشُرKashmiri
ᱥᱟᱱᱛᱟᱲᱤSantali
सिन्धीSindhi
कोंकणीKonkani
डोगरीDogri
ꯃꯩꯇꯩꯂꯣꯟManipuri
बरʼBodo
संस्कृतम्Sanskrit
অসমীয়াAssamese
मैथिलीMaithili
کٲشُرKashmiri
ᱥᱟᱱᱛᱟᱲᱤSantali
सिन्धीSindhi
कोंकणीKonkani
डोगरीDogri
ꯃꯩꯇꯩꯂꯣꯟManipuri
बरʼBodo
संस्कृतम्Sanskrit
অসমীয়াAssamese
मैथिलीMaithili
کٲشُرKashmiri
ᱥᱟᱱᱛᱟᱲᱤSantali
सिन्धीSindhi
कोंकणीKonkani
डोगरीDogri
ꯃꯩꯇꯩꯂꯣꯟManipuri
बरʼBodo
संस्कृतम्Sanskrit
অসমীয়াAssamese
मैथिलीMaithili
کٲشُرKashmiri
ᱥᱟᱱᱛᱟᱲᱤSantali
सिन्धीSindhi
कोंकणीKonkani
डोगरीDogri
ꯃꯩꯇꯩꯂꯣꯟManipuri
बरʼBodo
संस्कृतम्Sanskrit
09——FAQcommon questions
Questions
Everything you need to know about our data, evaluation, and annotation services.
10——Get Startedlet's talk
Better data. Better models.
Tell us what you need. Indic speech data, annotation, or model evaluation. We'll get back within 24 hours.