2,000 speakers of Marathi 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 Marathi specifically: Retains the retroflex lateral ळ, which has no Hindi or English equivalent and is frequently substituted with ल by non-native transcribers
• Code-mixing handled explicitly rather than edited out — Mumbai and Pune speech mixes Marathi, Hindi, and English in the same sentence. Marathi-only recordings collected in Pune under-represent the Mumbai reality of tri-lingual switching.

Sixteen to twenty-two weeks, run in waves so that recruitment in the harder catchments starts early rather than being left to the end.

Ten to twelve batches with rolling demographic reconciliation, since a quota gap discovered in week eighteen is very expensive to close.

A representative Marathi cohort should be split roughly 40% western Maharashtra, 25% Vidarbha, 20% Marathwada, 15% Konkan rather than concentrated in Pune.

At 2,000 speakers 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.

• Retains the retroflex lateral ळ, which has no Hindi or English equivalent and is frequently substituted with ल by non-native transcribers
• Mumbai and Pune speech mixes Marathi, Hindi, and English in the same sentence. Marathi-only recordings collected in Pune under-represent the Mumbai reality of tri-lingual switching.
• Available Marathi speech data is dominated by standard Puneri read speech. Vidarbha, Marathwada, and coastal Konkan varieties are severely under-collected, which is exactly where deployed voice products lose accuracy.

• 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
• ळ vs ल substitution by Hindi-trained transcribers
• Anusvara placement varies between conservative and modern orthography

100% technical QA, 7% content QA stratified across every axis, blind re-transcription audit, and per-cell demographic sign-off before final acceptance.

• 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: Long-tail demographic cells, which dominate both cost and schedule at this size.

• The final 15% of the cohort — the rarest demographic cells — routinely takes as long as the first half and should be scheduled first, not last
• Speaker pool exhaustion is a hard constraint in all but the largest languages
• Voice-based duplicate detection across 2,000 speakers is a genuine engineering task, not a spreadsheet check
• Marathi carries 6 recognised varieties across Maharashtra, Goa, parts of Karnataka, so the quota matrix is wider than the headline volume suggests
• Available Marathi speech data is dominated by standard Puneri read speech. Vidarbha, Marathwada, and coastal Konkan varieties are severely under-collected, which is exactly where deployed voice products lose accuracy.

• 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.