Industry · Voice AI, contact centres, IVR, speech analytics
Speech data for voice AI that has to work on a Pakistani phone line
Voice agents, IVR and call analytics fail in South Asia for boring reasons: 8 kHz telephone audio, Urdu–English code-switching every few words, budget handsets on weak networks, and accents the training set never heard. We collect exactly that audio — consented, transcribed to a written convention, balanced to your quota.
Who this is for
Built for teams shipping voice products to 255 million people, or to their diaspora
The same data serves a Karachi bank's IVR and a UK council's Urdu language line.
- Contact-centre AI vendors — ASR, agent assist, QA scoring and summarisation on real Urdu, Punjabi and Pashto calls
- Banks, telcos and utilities — self-service IVR and voice agents for Pakistan's 190 million mobile subscribers
- Speech platform providers — extending Whisper-class or in-house models to South Asian languages with production-distribution data
- Diaspora language services — NHS, council, telecom and healthcare lines serving Urdu and Punjabi speakers in the UK, Canada, US and Australia
- Speech analytics and compliance — call-outcome, intent and sentiment labels on channel-separated agent/caller audio
What breaks
Why models that score well on benchmarks fail on real calls
Open corpora are read speech in quiet rooms. Production is none of that.
Telephone band
Most South Asian voice traffic is 8 kHz GSM or VoIP with codec artefacts. Wideband-trained models lose fricatives and confuse retroflex consonants. Training on real telephony audio is the only fix.
Code-switching
Urban speakers switch into English mid-sentence — product names, numbers, verbs. Monolingual Urdu ASR drops half the content. Transcripts need a written convention for language spans and Roman-script tokens.
Accent and L1 spread
Urdu spoken by Punjabi, Pashto and Sindhi first-language speakers differs audibly. A model trained on Karachi Urdu degrades in Peshawar. Quota by L1 background, not just by language.
Acoustic reality
Open-air markets, traffic, generators, shared handsets on speaker. Studio-clean corpora teach the model an environment your callers never occupy.
What we supply
Speech data across the whole voice stack
Quota-controlled field collection in six cities and rural districts, transcription by native speakers, double-blind QA.
Telephony corpora
8 kHz channel-separated agent/caller audio in Urdu, Punjabi, Sindhi and Pashto, GSM and VoIP codecs, with intent and outcome labels and PII redaction. See the telephony corpus.
Conversational ASR data
Spontaneous two-party conversation, verbatim transcripts, diarization, code-switch tags, four age bands, urban/rural. See Conversational Urdu Speech.
Wake words and commands
Positive triggers at 0.3–5 m with graded noise, hard negatives and confusables in Urdu, Punjabi, Pashto and English. See the Wake Word & Command Set.
Accented English and TTS
Pakistani-accented English graded by CEFR and tagged by L1 for global contact centres; studio TTS voices in Urdu and Punjabi with pronunciation lexicons. See Pakistani-Accented English.
Ready-made
Catalog datasets for voice AI
Each ships with a datasheet, per-record consent artefacts and a free unwatermarked sample pack.
Pakistani Telephony Corpus
8 kHz call-center audio, four languages
SpeechConversational Urdu Speech
Spontaneous, diarized, consented
SpeechPakistani-Accented English
L1-tagged, CEFR-graded
SpeechWake Word & Command Set
Urdu, Punjabi, Pashto, English
SpeechRegional Languages Speech Pack
Saraiki, Balochi, Brahui, Hindko
AnnotationAnnotation on your calls
Transcription, intent and QA labels on audio you already own
How a programme runs
From your failure mode to a QA-passed corpus
Most programmes go from first message to first accepted batch in about six weeks.
Name the failure
Which calls fail: language, accent, band, noise, code-switching. That defines the quota and the acoustic spec.
Free sample, then pilot
Evaluate a sample pack against your own model in a day. A 50–200 hour pilot validates spec and convention in 3–5 weeks.
Weekly QA-passed batches
Delivered to your bucket with manifests, QA report and datasheet; out-of-tolerance audio re-collected free.
Hold-out test set
A held-out set from the same quota so you can measure WER on your users, not on a benchmark.
Answers
Questions voice AI teams ask
Meaningful gains typically start at 200–500 hours of in-domain telephony audio on top of a multilingual base model, and keep improving to a few thousand. Our guide on how much data an ASR model needs sets out the curve; a pilot on your own model is the real answer.
Yes. Native-speaker transcription with a written code-switching convention, diarization, intent, outcome and sentiment labels, and PII redaction, working remotely, in your VPC or on-prem. See Annotation & Evaluation.
Yes — read and spontaneous English graded by CEFR band and tagged by first language (Urdu, Punjabi, Pashto, Sindhi and more), which is what English-language IVR and analytics need for South Asian callers and agents.
Telephony corpora are collected from consenting contributors under a plain-language consent covering training use; we do not scrape or resell operator recordings. Consent artefacts travel with every record.
Yes — most of our licensing goes to AI teams outside Pakistan. English-language MSA and DPA, USD invoicing by default, W-8BEN-E for US payers, delivery to your AWS, GCP or Azure region. See Where we deliver for compliance mapping by region.
Which calls is your model failing on?
Tell us the language, the band and the failure mode. You get a collection design, a free sample and an honest view of what data will and will not fix.
Average first response: under 6 business hours. NDAs signed same day.