2,000 hours of Hindi Spontaneous Speech

Speakers talk unscripted on prompted topics — describing a process, recounting an event, arguing a position — with a moderator who prompts but does not lead.

A 60-minute session covering six to eight topics, recorded continuously and segmented afterwards into utterances at natural pause boundaries.

Yield at this style: 60–70% of recorded time is accepted. Long silences, moderator speech and abandoned topics are cut in post, so recorded hours run well ahead of delivered hours. Budget for the gap.

• Topic banks graded by familiarity, so speakers across education and occupation bands all have something to say
• Open prompts only — anything answerable with yes or no produces thirty seconds of audio and a stalled session
• Culturally grounded topics per region, since a prompt that works in Mumbai can draw blank looks in Guwahati
• Topic rotation across the cohort so the corpus does not over-represent a handful of subjects
• 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.

Screening is for willingness to talk, not literacy, which opens the cohort to speakers a scripted protocol would exclude. Expect to over-recruit by 15% for speakers who freeze on the day.

• 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.

• Segment boundary review — bad boundaries produce truncated words that poison training more than they help
• Disfluency transcription consistency audited across transcribers, since conventions drift fast on this style
• Moderator speech confirmed absent from delivered segments
• Topic distribution checked per speaker so one dominant subject does not skew the language model
• 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.

• Speakers who dry up after a minute, leaving sessions that look complete by duration but are mostly silence
• Drift into a reading register when a speaker becomes self-conscious about the microphone
• Moderator over-prompting, which turns a monologue corpus into an interview corpus
• Transcriber normalisation — quietly cleaning up disfluencies destroys the exact signal this style exists to capture

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
• Filled-pause and false-start markers
• Segment offsets into the parent session file
• Topic label per segment
• Speech-rate and pause-density statistics per speaker
• 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

• Robust ASR for real user speech
• Language modelling on natural syntax
• Disfluency detection and removal
• Prosody and speech-rate modelling

Phonetic balance cannot be guaranteed — speakers say what they say. If you need specific phone contexts or a controlled lexicon, this style has to be paired with a scripted layer.