Abstract
A realtime conversational voice LLM tends to mirror the language of its conversation history and of the caller's most recent utterance. When such an agent is reached through a preceding DTMF interactive-voice-response (IVR) menu in which the caller has explicitly selected a language ("press 2 for French"), two structural facts conspire to lose that selection: the IVR subsystem and the realtime-voice subsystem are two disjoint processes joined only by a stateless HTTP redirect, and the realtime model's mirroring tendency silently overrides a one-time greeting in the chosen language. This publication discloses a mechanism with two cooperating parts. First, a language carrier records the IVR-selected language at the instant of DTMF selection, keyed by caller identity, and threads it across the stateless redirect as a query parameter backed by a short-lived, single-consumption server-side map so it cannot leak to a later call. Second, at realtime-session-open the system resolves the effective language by a defined four-source precedence (IVR selection → returning-caller stored preference → persona default → platform default) and injects into the model session an imperative, persistent, anti-mirroring single-language instruction — a directive of the form "speak ONLY in {language} for the entire conversation; do NOT switch even if previous conversation history is in another language." The instruction is characterized by being (a) part of the durable session instructions rather than a one-shot greeting, and (b) explicitly anti-mirroring — it names the model behavior it suppresses. We present architecture, sequence, state-machine and data-model diagrams; a clean-room Node.js reference implementation; a prior-art delta against the closest references (US 9,667,785; US 9,401,993; US 9,654,631; US 12,293,161); an evaluation methodology; and an enumerated list of fifteen inventive aspects.
Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.
Recommended Citation
Assuncao, gustavo matthew, "IVR-to-Realtime Anti-Language-Drift Session-Instruction Injection", Technical Disclosure Commons, (July 13, 2026)
https://www.tdcommons.org/dpubs_series/10867