Skip to content
AIverseOne
All work
Government · Applied AILive in production

Sabha Record: the House, transcribed as it speaks

Telugu-first live voice-to-text, editorial correction, AI briefing and multilingual translation for legislative proceedings — running during the session, not after it.

ClientAndhra Pradesh Legislature
SectorGovernment · Record & proceedings
CoverageAssembly, secretariat, committees
StatusLive in production
What is Sabha Record?

Sabha Record is a voice-AI platform that turns live legislative proceedings, spoken in Telugu and English, into an accurate record, with briefs, summaries and translation.

The goal

Turn live spoken proceedings into an accurate, usable record while the session is still running.

What was in the way

Traditional transcription took days, and speakers switch between Telugu and English mid-sentence, with specialist vocabulary throughout.

What we built

A multi-model voice AI pipeline with live speech-to-text, editorial review, AI briefs and summaries, and translation, with people approving every record.

Records available during the session, not days later
Multilingual access to proceedings
Human sign-off keeps every record trustworthy
Voice AITranslationAI summariesHuman-in-the-loop
The pipeline
Multilingual voice-to-record pipelineChamber audio is diarised and transcribed, passes a human editorial gate, then fans out to briefs, summaries and translation before entering the official record. Four tasks are routed to four separate models.DURING SESSIONChamber audioOverlapping speakers,variable qualityDiarisationWho is speaking,and whenTelugu-first ASRCode-switching andprocedural vocabularyHUMAN GATEEditorial sign-offThe AI drafts.A person signs off.OfficialrecordCorrectable, citedSession briefPer sittingSummaryPoints of order, decisionsTranslationAcross languagesMULTI-MODEL ROUTINGOne model doing four jobs does all four badly.Each task goes to the model suited to it.TranscribeDEDICATED MODELCorrectDEDICATED MODELSummariseDEDICATED MODELTranslateDEDICATED MODEL
The editorial gate is the design decision that made the system acceptable for official use — nothing reaches the record without a person signing it off.
Context

Legislative proceedings must be captured verbatim and turned into a usable record. Traditionally this is stenographers, transcription pools, and a turnaround measured in days — during which the people who most need the record are working from memory.

The problem

Automatic speech recognition in a legislative chamber is close to the hardest version of the problem. Speakers overlap and interrupt. Members code-switch between Telugu and English inside a single sentence. The vocabulary is procedural and proper-noun-heavy: bill titles, constituency names, member names, points of order. Chamber audio is variable. And unlike a meeting transcript, the output becomes part of an official record — so an error is not an inconvenience, it is a correction that has to be made on the record.

What we built
  • 01

    Live Telugu-first transcription

    Assembly, secretariat and committee sittings transcribed during the session rather than after it.

  • 02

    An editorial layer

    The AI drafts, a person signs off. This is the design decision that made the system acceptable for official use.

  • 03

    Automated briefs and summaries

    Generated per sitting, so an official can read what happened without reading everything that was said.

  • 04

    Multilingual translation

    Proceedings become readable outside the primary language of the House, without a separate translation cycle.

  • 05

    A multi-model agentic pipeline

    Each task — transcribe, correct, summarise, translate — is routed to the model best suited to it, instead of forcing one model to do all four badly.

Why this is the hardest thing we have built

Anyone can integrate a speech API. Very few teams have taken Telugu ASR into a legislative chamber, added an editorial workflow that officials will actually sign off on, and kept it running session after session with the output entering the official record.