Indian Language Data for ASR Model Training

• Hundreds to thousands of hours of verbatim-transcribed speech
• Wide speaker diversity: age, gender, region, education, recording condition
• Speaker-disjoint train/dev/test splits

• Word error rate overall and per dialect
• Entity error rate on names and numbers
• Code-switch token accuracy

• Read-speech-only corpora that do not transfer to spontaneous audio
• Speaker leakage across splits inflating reported accuracy
• Normalised-only transcripts with the raw text discarded

Each of these is a defect that only becomes visible after training, when re-collection costs a release cycle.

A asr model training programme starts from the metric you need to move, not from an hour count. We work backwards: target metric, evaluation set design, then the training volume and speaker spread needed to reach it.

That means the evaluation set is specified and collected first, from speakers who never appear in the training data.

• 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.