Punjabi AI Training Data Collection

Punjabi 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.

• Punjabi is tonal: high, low and level tones distinguish words, and tone is not marked in Gurmukhi orthography
• Historical voiced aspirates surface as tone rather than aspiration, which confounds shared Indic phone sets
• Rural Malwai speech has distinctive vowel quality relative to Amritsar standard

Punjabi has 5 varieties that matter for data collection. Collecting only the standard variety produces a model that works in the capital and fails everywhere else.

• Majhi (standard)
• Malwai
• Doabi
• Puadhi
• Pothohari-influenced

Punjabi speech mixes Hindi and English freely, with strong diaspora influence in urban registers.

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

Tonal variation is essentially unmodelled in public Punjabi data, and Malwai/Doabi rural speech is scarce.

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 Punjabi-specific decisions that go into the style guide before any transcriber starts work.

• Tone is unrepresented in text, so pronunciation lexicons must be built from audio, not from spelling
• Shahmukhi vs Gurmukhi script decisions must be fixed per project
• Adhak (gemination) applied inconsistently

• Minimal-pair tone elicitation sets
• Agricultural and rural-finance domain conversation
• Spontaneous conversation across Majha, Malwa and Doaba

Cover all three historic regions (Majha, Malwa, Doaba); tone realisation differs measurably between them.

Punjabi sessions run in Amritsar, Ludhiana, Jalandhar, Chandigarh, Patiala. 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.