Hindi AI Training Data Collection

Hindi is not a variant of a language your model already handles. The specific properties below are the ones that show up as errors in production.

• Four-way stop contrast (voiced/voiceless x aspirated/unaspirated) that collapses in models trained on English-first acoustic units
• Retroflex series ट ठ ड ढ ण routinely mis-mapped to alveolar /t/ /d/ by imported lexicons
• Nasalisation (chandrabindu) is phonemic and is dropped by most off-the-shelf G2P front-ends
• Schwa deletion in word-final position varies by region, so the same orthographic word yields different pronunciations

Hindi has 7 varieties that matter for data collection. Collecting only the standard variety produces a model that works in the capital and fails everywhere else.

• Khari Boli
• Awadhi
• Braj
• Bhojpuri-influenced Hindi
• Haryanvi
• Bundeli
• Marwari-influenced Hindi

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.

Specify your expected English-token ratio in the brief. It is cheaper to collect the right mix than to filter the wrong one afterwards.

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.

This is the practical reason to commission collection rather than assemble open corpora: the gap in the public data is exactly the part your users occupy.

These are the Hindi-specific decisions that go into the style guide before any transcriber starts work.

• Inconsistent Devanagari vs romanised spelling for the same English loan word
• Nukta characters (क़ ख़ ग़ ज़ फ़) applied inconsistently by transcribers
• Numerals: whether to write digits, Devanagari numerals, or spelled-out words must be fixed in the style guide up front
• Honorific verb forms create long agreement chains that annotators shorten unless the guide forbids it

• Scripted prompt reading (phonetically balanced sentence sets)
• Wake-word and command-and-control utterances
• Digit strings, dates, amounts, and Indian address formats
• Two-party spontaneous conversation on everyday topics
• Simulated call-centre calls: billing, delivery, recharge, banking

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.

Hindi sessions run in Delhi, Lucknow, Jaipur, Patna, Bhopal, Indore, Varanasi. Multi-city fielding is the default because single-city cohorts carry an audible accent signature.

• 100% of delivered files pass automated technical QA for SNR, clipping, duration and silence
• 5-25% of files pass a second native-speaker content review, stratified by city, dialect and transcriber, and escalating to 100% on any batch that fails the agreed threshold
• Accepted yield runs 85-90% for scripted speech, 60-70% for spontaneous, 55-65% for conversational and 50-60% for telephony
• Default cohort quotas: 50/50 gender, with age bands at 30% (18-25), 40% (26-40) and 30% (41-60)
• 48 kHz / 24-bit capture, delivered as 16-bit PCM WAV, with studio sessions held below a -50 dBFS noise floor
• First response within one working day; a scoped, fixed quote within two to three

These are the figures a delivery is measured against, not aspirations. A batch that misses them is re-recorded at our cost rather than repaired.