Standard Automatic Speech Recognition (ASR) models are remarkably proficient at transcribing common language, but they often stumble on "long-tail" terms—specific contact names, niche song titles, or technical jargon. To solve this, researchers use contextual biasing, a technique that provides the model with a "cheat sheet" of relevant phrases (a bias list) to look for in the audio.
However, providing a cheat sheet introduces a new problem: the model may become "trigger-happy," hallucinating rare words from the list even when they weren't actually spoken. Recent research focuses on teaching models not just how to use these lists, but crucially, when to ignore them.
The Mechanism of Deep Biasing
In modern end-to-end ASR, contextual biasing is often implemented as a "deep biasing" layer. This layer attends to an external list of phrases and injects that information directly into the model's internal representations.
In Contextualized End-to-End Speech Recognition with Contextual Phrase Prediction Network, Kaixun Huang and colleagues describe a Contextual Phrase Prediction (CPP) network. This acts as an auxiliary task: while the model transcribes audio, it simultaneously predicts which phrases from the bias list are present. By providing explicit supervision for these rare words, the researchers achieved a 40.5% improvement in recognizing targeted phrases.
The Hallucination Trap
The danger of deep biasing is over-reliance. If a model is always trained with a perfectly matching bias list, it learns that if a word is on the list, it is likely in the audio. In a real-world deployment, the bias list might contain hundreds of names, only one of which (or none at all) is relevant to the current utterance. Without proper training, the model may "force" a match, leading to a high false-positive rate.
To build intuition, imagine a student taking a multiple-choice test who is told that the answer is always one of three specific words. If the actual answer is something else entirely, the student will still pick one of those three, resulting in an error.
Teaching the Model to Say "None of the Above"
To mitigate this, researchers have introduced mechanisms that force the model to evaluate the relevance of the bias list.
1. The Null Token
Kaixun Huang and his team addressed this by including a dedicated null token within the external context list. This allows the CPP network to explicitly signal when no relevant phrases are present in the audio. By making "no match" a valid and supervised output, the model learns to distinguish between a genuine acoustic match and mere presence in the bias list.
2. Selective Deactivation
In Text Injection for Neural Contextual Biasing, Zhong Meng, Rohit Prabhavalkar, and their team at Google explored Contextual Text Injection (CTI). Their approach uses massive amounts of unpaired text to simulate speech-like representations for rare words.
Crucially, they found that to prevent hallucination, the biasing mechanism must be selectively deactivated during training. By frequently providing the model with empty bias lists, the model is forced to rely on its standard recognition capabilities. This "dropout-like" strategy for context ensures that the biasing layer acts as a helpful hint rather than a crutch, preserving general accuracy while still allowing for a 43.3% reduction in error rates for rare words.
Why It Works
These methods work by breaking the false correlation between the presence of a word in a bias list and its existence in the audio. By incorporating "negative examples"—scenarios where the bias list is empty or irrelevant—the model learns a more robust gatekeeping mechanism. It only allows the contextual information to influence the final transcription when there is strong evidence in the acoustic signal, balancing specialized performance with general-purpose reliability.