Custom Vocabulary and Hotword Injection in Commercial ASR APIs

Vendors offer two ways to inject domain vocabulary into speech recognition systems.

Cover illustration for “Custom Vocabulary and Hotword Injection in Commercial ASR APIs”
Written by
The Voice Layer Editors
Published
October 2, 2026
Reading time
9 min read
Sources cited
7 sources ↓

If a transcription tool outputs "very site" instead of "Veracyte", it isn't malfunctioning.

Why general-purpose ASR fails on domain vocabulary

Veracyte, a company name, is heard as "very site." Nexlify, a product, becomes "next-fi." KRATOS, an internal system, ends up as "NATO." The audio is not unclear, and the microphone is not bad. Modern speech recognition is probabilistic: when decoding, the model chooses the word sequence it considers most likely based on its training. Uncommon product, drug, and proprietary acronym terms show up so rarely in that data that the model gives them almost no chance of being right. Deepgram quantifies the result: roughly 40% of medical terms get missed by general-purpose recognition systems for precisely this reason.

The frequent word prevails because the model’s prior favors it, though the signal can point equally well toward the rare term. It is making a sound statistical choice, just not for the domain it now finds itself in. Since transcription typically comes first in a pipeline, one wrong word there can ripple through all later stages.

Commercial providers do not rebuild the entire acoustic model separately for every customer. In practice, providers keep the model's weights unchanged and adjust decoding during transcription, supplying a prior list of terms to favor for the work at hand.

Shallow Fusion, Neural Biasing, and the Boost Parameter

With word boosting, decoding is steered by adjusted scoring rules, not by relearning the model’s parameters. In production systems, this is usually handled through shallow fusion in the speech-recognition sense: score = acoustic_score + λ × language_model_score, plus a vocabulary-bias component scaled by β. The acoustic score shows how well a possible term lines up with the sound. The β part supplies the boost: its weighted adjustment tilts scoring toward named terms, with β setting the strength of that pull.

The boost has an effect only when the audio has already put the listed term within reach. When the model is weighing "next-fi" against "Nexlify," because the audio could reasonably support either, a boost nudges the choice in favor of "Nexlify". A boost adjusts which live candidates come out on top. Terms absent from the candidate set remain absent, and the model's weights stay the same, leaving the system no more or less capable after biasing than before.

These approaches serve distinct purposes rather than being mere alternatives; each is built for specific scenarios, so selecting one represents the initial choice any developer faces. Neural biasing in CLAS-style systems operates at a more fundamental level: instead of applying a post-hoc score adjustment, it allows the model to directly access an embedded phrase list while decoding, creating a more nuanced understanding of the target term's acoustic properties.

Static pre-compiled lists vs. runtime context injection

Vocabulary biasing in commercial ASR takes two operationally distinct forms, neither of which is merely the same feature in different packaging. Each is meant for a different situation, and the developer's first genuine call, made before any integration code gets written, is which one to go with.

One approach uses a static list prepared beforehand. The terms are collected ahead of time by a developer, kept by the provider, and used in recognition through an ID reference instead of transmitting the vocabulary. As audio is interpreted, listed phrases are favored by the model, while each request carries just the ID rather than the stored terms. For example, Alibaba Cloud tells users a list change may not become active until five minutes or more have passed. That bargain works well when the vocabulary is stable, but it becomes clumsy when the word set shifts minute by minute.

The other approach, runtime context injection, is also called instant hotwords, or alternatively direct-pass. Nothing gets registered ahead of time here. The terms ride inside each request, apply only during that one session or call, and need no list ID. It fits best where the relevant wording cannot be known before a session begins: who is joining a meeting, which customer account a support call concerns, or whatever phrases belong to that one exchange alone. To take one case, Volc, the streaming ASR API from ByteDance, accepts hotwords in each request, though it limits both the count of terms and their combined text, with a ceiling of 300 characters.

That fits conversations that evolve with time rather than vocabulary captured in a fixed list of terms.

Whether to use static lists or runtime injection is not simply a question of preference. It follows from one operational question: are the words needing a boost available before the session starts, or do they show up only after it has begun? Durable, reusable terms belong on the static side. Words tied to one caller, one meeting, or one account should be injected at runtime.

How the major commercial APIs implement these mechanisms

Cloud Speech-to-Text builds its system on structured phrase sets and a boost value the developer can set between 1 to 20, and the documentation spells out the trade-off: raise the boost too far and false positives appear.

AWS Transcribe follows another pattern. Instead of injecting terms with each request, users point AWS Transcribe to uploaded custom vocabulary files treated as standing resources, while custom language models handle domain terminology and the service includes diarization plus a medical-only transcription endpoint.

With Deepgram, self-serve keyword boosting can go directly into a request, keywords=Nexlify:2 for instance, and a developer gets custom vocabulary too, with no data science background needed to configure.

Alibaba Cloud Model Studio, spanning its Fun-ASR, Paraformer, and Qwen-Audio model lines, is the most complete of the group in terms of mechanism coverage: it supports precompiled hotwords by list ID, instant hotwords inline per request, and context enhancement through conversation history, all three of the modes described above. Hotword features work only in the primary workspace, and every sub-workspace is excluded from them, including Singapore sub-workspaces, which are specifically named among the exclusions.

ByteDance's Volc streaming ASR accepts hotwords directly in its request payload, structured under request.corpus.context using {"hotwords":[{"word":"…"}]}, with list size controlled and total text held to a 300-character ceiling, and developers can use boosting_table_id when working from a pre-compiled list.

Calibrating boost values against over-biasing

Pushing a boost value high enough makes the model go beyond preferring the listed term and begin adding it to transcripts even when no such word was ever spoken, producing an invented entry that harms the output more than the initial error did. Over-biasing, as this failure mode is called, appears in the literature from the very first studies of contextual biasing onward, and developers accept it as their primary hazard whenever they tune boost values.

No engineering trick eliminates this trade-off; it can only be handled with care. A biasing setup must lean firmly enough that the model recognizes the term whenever it is genuinely present in the recording, without leaning so firmly that the term crops up where nobody said it. Each platform finds that equilibrium in a different place. AWS Transcribe offers no adjustable boost setting, instead letting the order of terms within its vocabulary file determine priority.

The mechanism's behavior points to a few practical rules. Use the platform’s baseline, whether default or minimum boost, and increase it only when one particular term is repeatedly heard wrong even though its sound matches the clip closely enough for boosting to help. Check the boosted setup against clips that lack the target term, since that screening is the quickest way to spot hallucination before release. Choose a small, selective set of precise terms instead of an expansive catch-all list, because the next section shows how simply adding more entries can hurt recognition accuracy. Also keep in mind boosting’s limits: it will not overcome muffled recordings, loud interference, or speech patterns too distant from the model’s training data, because it merely reorders options the model already treats as plausible.

Rev.ai uses phrase rules to show calibration can be built into the API's design rather than left all to developer judgment. Such caps reduce the space in which excess biasing might occur.

Scale, interference, and retrieval-based pre-filtering

Another major way vocabulary biasing breaks down is unrelated to the strength applied when boosting an individual term. It comes down to how many terms sit on the list at the same time. Larger hotword lists cause entries to clash, driving up mixed error rates across retrieval-based and fusion-based biasing methods alike. This represents a fundamental change in behavior beyond a specific list threshold, not a gradual erosion of accuracy as more entries are added.

Two 2025 studies converge on the same answer: cut down the list before biasing starts, rather than making the biasing mechanism handle everything together. In Alibaba's contextual-biasing study, the proposed fix is RADA, short for Robustness-Aware Data Augmentation. Biasing works better against a shorter set focused on terms that still trip up the model, instead of an overstuffed list of items it already handles.

Kuaishou Technology's researchers took a similar path from another angle with H-PRM, unveiled in a CIKM '25 paper. Benchmarks comparing H-PRM with SeACo-Paraformer plus speech-capable LLMs such as Qwen2-Audio-Instruct and Whisper-small showed a 29.44% mean drop in mixed error rate over four datasets. BR-ASR, a kindred technique unveiled in an Interspeech 2025 paper, pursues that aim via retrieval trained with contrastive learning, sifting out pertinent bias words before the ASR model ever sees them.

These research efforts speak directly to the business constraints programmers face right now. The work happening on retrieval layers targets the root cause of those caps, while maturing filtering done before retrieval offers the clearest route to commercial APIs that can handle longer vocabulary lists without taking the accuracy hit that forces teams to keep those lists short today.

When prompt-based and structured biasing diverge

Some providers draw a real line between hotword lists and natural-language prompts rather than treating them as interchangeable interfaces. AssemblyAI describes the keyterms feature and free-form context prompt as tools designed to work together, not stand as two options where a developer must pick one. With a structured list, the aim is exactness: the decoding process gets a set group of terms, and can give each one its own weight. With a prompt, the aim is context: the text frames the setting, exchange type, or topic, and the model uses that to adjust its wider expectations rather than single out specific terms for extra weight.

That difference reflects a bigger split that cuts across the entire terrain. Structured lists of either kind, whether drawn up beforehand or supplied on the fly, shine when the terms involved are known ahead of time as a closed inventory of concrete items: drugs, brands, named entities, acronyms. On the other side, prompting and context enhancement come into their own when the aim is not to pin down a set vocabulary but to adjust the model's overall feel for the dialogue it is hearing: how formal it sounds, what domain it belongs to, how the back-and-forth unfolds. The context enhancement mode built by Alibaba, made for exchanges that develop as they unfold rather than for a preset term inventory, falls on that side of the divide. The instruction-driven prompting that LLM-native tools such as Fun-ASR and Index-ASR rely on sits there as well, since their hotword lists travel inside a text prompt rather than reaching a biasing module through some dedicated structured channel.

Picking a structured list, a prompt, or both as AssemblyAI's documentation recommends hinges on the same point separating static lists from runtime injection in the first place: is the fix about specific nameable words, or a wider shift in what language the model should anticipate?

Methodology & sources

  1. Improve recognition accuracy-Alibaba Cloud Model Studio(Model Studio)-阿里云帮助中心

    Provided details on Alibaba Cloud's context enhancement mode and its distinction from preset term inventories.

  2. H-PRM: A Pluggable Hotword Pre-Retrieval Module for Various

    Source for H-PRM research from Kuaishou Technology, including the 29.44% mean drop in mixed error rate benchmark and the finding that larger hotword lists cause interference across biasing methods.

  3. feat: ASR 热词注入(Volc / Qwen3-ASR) by IchenDEV · Pull Request #98 · IchenDEV/utter

    Provided implementation details on Volc streaming ASR hotword injection, including the request.corpus.context structure and 300-character ceiling.

  4. Overview

    Provided information on Rev.ai's phrase rules approach, illustrating how calibration can be built into an API's design.

  5. Improve recognition accuracy - Alibaba Cloud Model Studio - Alibaba Cloud Documentation Center

    Provided details on Alibaba Cloud Model Studio's hotword mechanisms, including the five-minute activation delay, list ID usage, instant hotwords, and sub-workspace exclusions.

  6. Fun-ASR Technical Report Tongyi Fun Team Alibaba Group

    Source for details on Fun-ASR's instruction-driven prompting approach where hotword lists travel inside a text prompt.

  7. Index-ASR Technical Report

    Source for Index-ASR's LLM-native prompting approach to hotword biasing.

The Voice Layer Editors