2,000 hours of Hindi Telephony Speech

Speech captured over a real telephony path — a genuine call through the network, not studio audio downsampled afterwards to imitate one.

Scenario-driven calls between a caller and an agent or IVR flow, each leg recorded separately, across a deliberate spread of handsets and network conditions.

Yield at this style: 50–60% of recorded time is accepted. Dropped calls, network artefacts beyond tolerance and unusable legs are discarded. Real network conditions are the point of the style and also its main cost.

• Call scenarios drawn from real contact-centre intents: balance enquiry, complaint, booking change, escalation
• IVR flows scripted with deliberate mis-entry and barge-in paths, since those are where deployed systems fail
• Handset spread specified up front — low-end Android, feature phone, landline — because handset variance is a real acoustic axis
• Background conditions varied on purpose: street, vehicle, indoor, since real callers are rarely in quiet rooms
• Built against Hindi specifically: Four-way stop contrast (voiced/voiceless x aspirated/unaspirated) that collapses in models trained on English-first acoustic units
• Code-mixing handled explicitly rather than edited out — Urban Hindi speech is Hinglish in practice. Expect 15-40% English tokens in spontaneous speech: numbers, brands, technology terms, and whole clause switches. Any Hindi corpus that excludes English tokens will not match production traffic.

Twenty to twenty-eight weeks, run in two or three waves. Wave one establishes the protocol and the transcription convention; later waves scale against a proven baseline rather than an assumed one.

Continuous fortnightly delivery from week four, roughly fourteen batches, each with acceptance testing and a rolling demographic reconciliation against the target quota matrix.

Largest recruitment pool in the network. A 1,000-speaker Hindi cohort with balanced gender and 18-45 age bands is typically fielded across four cities to avoid a single-city accent bias.

At 2,000 hours the quota matrix is enforced per cell, not in aggregate. Every dialect, age and gender combination carries its own target and is signed off individually before final acceptance, because an aggregate 50/50 split can hide a cell that was never filled at all.

Recruitment must be stratified by handset and carrier as well as by dialect, which adds a screening axis the other styles do not have.

• Four-way stop contrast (voiced/voiceless x aspirated/unaspirated) that collapses in models trained on English-first acoustic units
• Urban Hindi speech is Hinglish in practice. Expect 15-40% English tokens in spontaneous speech: numbers, brands, technology terms, and whole clause switches. Any Hindi corpus that excludes English tokens will not match production traffic.
• Public Hindi corpora skew heavily towards read newspaper text from educated urban speakers in Delhi and NCR. Rural Bihar and eastern UP speech, elderly speakers, and low-literacy speakers reading prompts aloud are largely absent.

• Codec and sample rate verified per file, rejecting any studio audio that has been downsampled to fake a telephony path
• Echo and double-talk checked on both legs
• DTMF events verified against the call log
• Level normalisation applied per leg, since handset output levels vary far more than studio microphones do
• Inconsistent Devanagari vs romanised spelling for the same English loan word
• Nukta characters (क़ ख़ ग़ ज़ फ़) applied inconsistently by transcribers

100% technical QA, 5% content QA stratified across every axis, a blind 2% re-transcription audit, and a monthly cross-wave consistency review comparing early and late batches for drift.

• Studio audio downsampled to 8 kHz and passed off as telephony, which has none of the codec or packet-loss characteristics that matter
• Echo and double-talk that make the agent leg unusable
• Carrier and handset monoculture, producing a corpus that only represents one acoustic path
• Over-clean recordings from participants who move somewhere quiet to take the call, defeating the purpose

Cost at this band is driven by: Wave structure and the programme governance it requires; Long-tail demographic and dialect quotas, which dominate the final third of the build.

• Drift between waves is the defining risk: audio recorded in month one and month six must be indistinguishable in convention, or the corpus splits into two datasets
• Speaker pool exhaustion is a live constraint in all but the largest languages and shapes which cities are used
• Staff turnover across twenty-eight weeks is a certainty, so handover documentation is part of the deliverable rather than an afterthought
• Storage, transfer and manifest integrity become engineering problems in their own right at this size
• Hindi carries 7 recognised varieties across Uttar Pradesh, Bihar, Madhya Pradesh, so the quota matrix is wider than the headline volume suggests
• Public Hindi corpora skew heavily towards read newspaper text from educated urban speakers in Delhi and NCR. Rural Bihar and eastern UP speech, elderly speakers, and low-literacy speakers reading prompts aloud are largely absent.

• WAV audio to your naming convention, with the per-file manifest
• Verbatim Devanagari transcripts with utterance-level timestamps
• Call metadata: duration, codec, handset class, carrier, packet-loss events
• Intent label per call and per turn
• DTMF and hold, transfer and barge-in events
• Agent and caller leg identifiers
• Per-speaker metadata: age band, gender, region, dialect, education band
• Consent records mapped to speaker IDs
• QA report with pass rates, rejection reasons and agreement statistics
• Speaker-disjoint train / dev / test splits on request

• Contact-centre and IVR ASR
• Voice bots operating over the phone network
• Intent classification on narrowband audio
• Robustness to codec and packet loss

Narrowband telephony audio is the wrong input for TTS or any wideband model — the frequency content simply is not there. Use it for models that will be deployed on a phone line and nothing else.