2,000 hours of Hinglish Scripted Speech

Speakers read prompts from a phonetically balanced script shown one line at a time, with retakes on any flubbed or misread line.

A 45-minute booth session yields roughly 300–400 prompts. One file per utterance, cut at the prompt boundary, so alignment is exact before any forced-alignment pass runs.

Yield at this style: 85–90% of recorded time is accepted. Prompt-level retakes catch problems inside the session, so very little is discarded afterwards. This is the highest-yield style we run.

• Script built for triphone coverage rather than word coverage, so rare phone contexts appear often enough to train on
• Digit strings, dates, currency and person names blocked separately, because these are where deployed ASR actually fails
• Domain lexicon injected from your product vocabulary when you supply one
• Sentence-length distribution spread deliberately, since all-short prompts produce a model that cannot handle long utterances
• Built against Hinglish specifically: Intra-sentential switching means English words carry Indian phonology, so English acoustic models mis-transcribe them
• Code-mixing handled explicitly rather than edited out — Hinglish is the code-mixing case itself. Typical urban customer-support speech is 30-60% English tokens embedded in Hindi grammar, with switching several times per utterance.

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 switching behaviour, not by language proficiency. Screening recordings are used to confirm speakers switch naturally rather than performing one language.

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.

Requires literate speakers comfortable reading aloud in the target script, which is the main constraint on cohort breadth and has to be actively counterweighted.

• Intra-sentential switching means English words carry Indian phonology, so English acoustic models mis-transcribe them
• Hinglish is the code-mixing case itself. Typical urban customer-support speech is 30-60% English tokens embedded in Hindi grammar, with switching several times per utterance.
• Almost no public corpus contains genuine intra-sentential Hindi-English switching with per-token language tags. This is the highest-value gap for anyone building Indian conversational AI.

• Prompt-to-audio match verified by a native reviewer; a misread line is a rejected file, not an edited transcript
• Hyperarticulation flagged — a speaker over-enunciating because they are reading produces audio that does not match deployment
• Clipping and truncation checked at both utterance boundaries
• Per-speaker prompt coverage confirmed, so no speaker silently skips a block
• Whether English tokens are written in Latin or transliterated into Devanagari must be fixed by rule, not left to annotators
• Language-ID tagging per token is required for training but is skipped by most vendors

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.

• Reading voice: flat prosody and unnatural stress that trains a model on speech nobody actually produces
• Prompt fatigue in the back half of long sessions, where accuracy drops and pace flattens
• Speakers who are not fluent readers, which quietly biases the cohort toward higher education bands
• Script leakage across speakers, producing a corpus that memorises sentences instead of covering sounds

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
• Hinglish carries 4 recognised varieties across Delhi NCR, Mumbai, Bengaluru, so the quota matrix is wider than the headline volume suggests
• Almost no public corpus contains genuine intra-sentential Hindi-English switching with per-token language tags. This is the highest-value gap for anyone building Indian conversational AI.

• WAV audio to your naming convention, with the per-file manifest
• Verbatim Devanagari + Latin transcripts with utterance-level timestamps
• Prompt ID mapped to every utterance
• Verbatim deviation flags where the speaker departed from the script
• Per-utterance SNR and duration in the manifest
• 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

• ASR acoustic model baselines
• TTS voice building where a single speaker is recorded at depth
• Pronunciation lexicon and G2P validation
• Forced-alignment and phone-boundary work

Scripted audio contains no disfluencies, no false starts and no natural turn-taking. A model trained on it alone degrades sharply on real spontaneous input, so it is a baseline layer rather than a complete training set.