A Language-Routed Speech-to-Action Framework for Code-Switched Indic Voice Assistants

There are 22 Scheduled Languages in India, along with code-switching, and accents in different dialects make voice the most viable medium for many users of digital public services. There have been remarkable strides in Automatic Speech Recognition (ASR) technology for Indic languages, but most assessments are still limited to word error rate (WER). Conversely, LLM agents rely on clean input text and thus conceal the compound failure when the spoken and code-switched object is mapped to the Application Programming Interface (API). In this paper, we introduce IndicAgent-ASR, a cascade of acoustic-to-agent system that performs language/code switch detection, routes speech to Indic or English ASR models, canonicalizes multilingual slots, plans for tool call or dependent tool call, runs a safety filter, and generates a spoken response using Indic text-to-speech. We specify the architecture, an end-to-end flowchart, a turn-level algorithm, and an easily reproducible software stack. Published benchmarking results include multilingual and code-switching baselines for MUCS 2021 with WER scores of 30.73% and 32.45%, best-performing accent models with WER ranging from 10.5% to 20.5% in the LAHAJA benchmark, and parameter filling accuracy of 60.6% (English) against 39.2% (Indic) in VoiceAgentBench using the strongest ASR-LLM pipeline, where sequential dependent calls drop to 10.0% and 4.3%. The feasibility prototype of the Hindi/Hinglish and Marathi language features mandi, weather, e-health record, and balance check tools, but denies voice fund transfer and prompt injection tools. The evidence supports keeping a strong Indic ASR front-end in the cascade and treating dependent tools and safety as first-class modules.