2,000 hours of Tamil Conversational Speech

Two speakers hold an unscripted conversation seeded with a scenario, recorded on separate channels so overlap and turn-taking survive into the delivered files.

A 45–60 minute paired session with both speakers on isolated microphones, either in adjacent treated rooms or split-mic in one room with bleed measured and logged.

Yield at this style: 55–65% of recorded time is accepted. Overlap regions, crosstalk bleed and one-sided stretches all cost delivered time. The lowest-yield studio style we run, and priced accordingly.

• Scenario seeds rather than scripts — a disagreement to resolve, a plan to make, an experience to compare
• Pairing designed deliberately: familiar pairs produce natural interruption, stranger pairs produce polite turn-taking, and you need both
• Scenarios that invite disagreement, because agreeable conversation produces almost no overlap to train on
• Register mixed across pairs so the corpus is not uniformly formal
• Built against Tamil specifically: Extreme diglossia: written Tamil and spoken Tamil differ so much that read-speech corpora are near-useless for conversational ASR
• Code-mixing handled explicitly rather than edited out — Tanglish is the default urban register. Technology, finance, and workplace vocabulary is largely English embedded in Tamil syntax.

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.

Recruit by district rather than by city alone; Chennai-only cohorts produce models that degrade sharply in the south and west of the state.

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.

Effectively doubles recruitment load, since speakers are booked in matched pairs and a single drop-out cancels the whole session. Plan on 20–25% over-recruitment.

• Extreme diglossia: written Tamil and spoken Tamil differ so much that read-speech corpora are near-useless for conversational ASR
• Tanglish is the default urban register. Technology, finance, and workplace vocabulary is largely English embedded in Tamil syntax.
• Almost all public Tamil audio is literary read speech from news or scripture. Genuine colloquial Tamil, especially southern and Kongu varieties, is the single biggest gap for anyone building Tamil voice products.

• Channel bleed measured on every session and rejected above the agreed threshold
• Diarisation labels verified against the isolated channels rather than inferred from the mix
• Turn boundaries and overlap regions checked by a native listener
• Speaking-time balance per pair audited, so a dominant speaker does not silently halve the session's value
• Transcribers normalising spoken Tamil into literary Tamil, destroying the acoustic-text alignment
• ழ / ள / ல confusion

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.

• Crosstalk bleed that makes clean per-speaker training targets impossible to recover afterwards
• One speaker dominating, leaving a pair that delivers half the expected audio
• Unnatural politeness between strangers, producing clean but unrepresentative turn-taking
• Scheduling attrition — both speakers have to show up, so no-show rates compound rather than add

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
• Tamil carries 6 recognised varieties across Tamil Nadu, Puducherry, parts of Karnataka and Kerala, so the quota matrix is wider than the headline volume suggests
• Almost all public Tamil audio is literary read speech from news or scripture. Genuine colloquial Tamil, especially southern and Kongu varieties, is the single biggest gap for anyone building Tamil voice products.

• WAV audio to your naming convention, with the per-file manifest
• Verbatim Tamil transcripts with utterance-level timestamps
• Speaker-turn segmentation with start and end timestamps
• Overlap regions marked with participating speaker IDs
• Backchannel and interruption markers
• Per-pair relationship metadata: familiar or stranger
• 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

• Speaker diarisation
• Meeting and multi-party ASR
• Turn-taking and endpointing for voice agents
• Speaker separation and target-speaker extraction

Two-party conversation does not generalise to multi-party meetings with four or more speakers, where overlap statistics change substantially. Specify that case separately.