Indian code-mixed calls have no single correct transcript: LLM, CRM and compliance each need a different one
BEYOND__7 · reddit · 2026-11-04
A developer building voice pipelines for Indian languages explains that code-mixed calls (e.g. Hinglish) have four equally 'correct' transcript forms: verbatim Devanagari, cleaned-up Devanagari with numerals, mixed-script, and all-Roman. Each downstream component wants a different one — LLMs work best on the cleaned version, keyword alerts only fire reliably on mixed script, CRMs need Roman, compliance needs verbatim, and TTS breaks on mixed scripts. Whisper's random script-switching on Hinglish makes conversions messy. Two approaches: transcribe once and convert downstream (AI4Bharat open-source transliteration models + number normalization), or use an STT with selectable output formats like gnani Prisma v3's four modes — though you must send audio twice if you need two formats.
More from coding & agent
- Does more model intelligence still matter for coding past a threshold? — the_juan_1 · 2026-11-05
- Get a real second opinion: send the failure, not the agent's theory — Ok_Negotiation_2587 · 2026-11-05
- Dev builds code-driven Poison Ivy scrolling component in Framer with multiple configs — evilrabbit_ · 2026-11-05
- The Missing Primitive in AI-Employee Tooling: Agents Can't Watch and Wait — corbtt · 2026-11-05
- rustgo: a from-scratch Rust compiler written in Go, mirroring rustc's pipeline — rakyll · 2026-11-05
- Atlas launches: a hypervised Mac desktop giving autonomous agents their own hidden workspace — Scobleizer · 2026-11-05