Patent · Speech Recognition

Indicator Loss for Context-Aware ASR

Making contextual speech recognition auditable — first inventor on patent CN113808593B.

Google Patents page for CN113808593B — Voice interaction system, related methods, devices and equipment — Granted patent CN113808593B (first inventor) — personalized end-to-end speech recognition for voice interaction.

Problem

Voice assistants need to bias recognition toward contacts, device names, and other personal vocabulary — a problem now often called dynamic vocabulary-based contextual biasing. The bias list changes per user or session and must be injected at runtime, without retraining the base ASR model.

Prior approaches (including Google’s CLAS-style contextual ASR) often treated biasing as an opaque soft attention over a list — hard to inspect, hard to update in real time, and hard to debug when the model ignored the bias.

Approach

I invented Indicator Loss: an explicit, step-level supervised gate that decides when the decoder should attend to the bias list. Instead of relying only on decoder loss to hope the model “notices” a key phrase, Indicator Loss trains a binary relevance signal over the bias phrases at each decoding step — so operators can see whether biasing fired, not only the final transcript.

This sits in the same family of ideas that later surveys describe as dynamic gating / activation and auxiliary bias-token supervision: mechanisms that selectively enable biasing rather than always soft-fusing the full list.

Industry comparison of personalized ASR decoders: traditional ASR, Google CLAS, and Indicator Loss architecture — Industry comparison: traditional ASR → Google CLAS → personalized decoder with Indicator Loss.

Deployed on Alibaba’s Tmall Genie at national scale, the design enabled real-time bias-list updates without weeks-long language-model retraining — the practical requirement behind dynamic vocabulary biasing in production assistants.

Indicator Loss results compared with Google CLAS across IoT and contact test sets — Head-to-head evaluation against CLAS on IoT and contact personalization sets.

Impact

Granted Chinese patent CN113808593B (first inventor; filed 2020, granted 2025)
Production deployment on a leading smart-speaker platform
An early production instance of what the field now frames as dynamic vocabulary / contextual biasing — with an inspectable gate rather than opaque attention alone
Independent academic work (Interspeech 2023–2025 and later surveys) continued to develop related themes: bias encoders, gating, and specialized bias objectives

References

Patent: CN113808593B
Survey / topic overview: Dynamic Vocabulary-Based Contextual Biasing (Emergent Mind)

← Back to home · All projects