Code-mixed speech, correct script
Premium थोड़ा ज़्यादा है but the cashless network बहुत बड़ा है।
One utterance, two languages, two scripts, and a product term that must stay in English. This is the normal case on an Indian sales call, and it is where most voice AI quietly fails.
The failure modes we test for
Wrong script
Hindi output rendered in Latin transliteration when the CRM, the SMS gateway or the compliance log expects Devanagari — or the reverse.
Over-transliteration
English product names, plan tiers and legal terms transliterated into Devanagari when they should remain in Latin script, making them unsearchable and sometimes unrecognisable.
Script drift mid-response
A response that starts in Devanagari and slides into Latin partway through, or mixes both within a single word.
Mangled named entities
Customer names, place names and policy numbers corrupted by transliteration round-trips.
Register collapse
Formal Hindi where the caller used colloquial speech, or the reverse — technically correct output that no agent would ever say.
Slang and regional variation
Vocabulary that appears in real calls and in almost no training corpus, including local product slang and abbreviations specific to your business.
The language model is not the weakest link
Speech recognition is the real bottleneck. Most ASR is trained monolingual, and word error rates degrade sharply on speech that switches language mid-clause. If transcription is wrong, nothing downstream can recover. We measure this first, before any model work.
Speech synthesis is the second bottleneck. A Hindi voice reading “premium” or “cashless network” needs correct English pronunciation, not Devanagari-phonetic. Most TTS handles one language well and code-mixed audio badly, and callers hear it within two seconds.
What we host and what we do not. We fine-tune and serve the language and compliance layer. For speech recognition and synthesis we benchmark options and tell you which are self-hostable at acceptable quality for your languages and which are vendor choices. We will not claim the whole pipeline is solved when it is not. [Fill in your current stack decisions here before launch.]
The cost argument, specific to Indic scripts
- Devanagari requires substantially more tokens than Latin script for equivalent meaning in standard BPE tokenisers.
- Per-token API pricing therefore penalises Hindi output relative to English, above the headline rate.
- Self-hosted inference carries no per-token cost, so the gap widens with volume rather than narrowing.
- We are measuring the exact multiplier on real transcripts. Until that is published we state the direction, not a number.