Keyterm Filtering for STT: Reducing Hallucinations in Non-Standard Accents

Speech recognition has been improving its overall accuracy for years, but there's one specific problem that remains unresolved: when you input custom keyterms—brand names, technical terms, industry jargon—to enhance transcription, STT models tend to "see" them even when they're not in the audio. The effect is especially pronounced in non-standard English or with non-normative accents.

That's exactly what Aditu is trying to solve, a project that appeared this week on Hacker News in the thread "Show HN: Keyterm Filtering for Voice AI". Its proposal is straightforward: a postprocessing filter layer that analyzes transcriptions generated with keyterm prompting and removes those terms that probably weren't in the original audio.

The Problem with Keyterm Prompting

Keyterm prompting is a standard technique in voice pipelines: you tell the STT engine which words it should prioritize or recognize, which helps with proper names, brands, or uncommon technical vocabulary. Providers like Deepgram implement it natively.

The downside is that the model, having these words "in mind," inserts them into the transcription even when the speaker didn't say them. With neutral American or British English accents, the problem is manageable. With Hindi, English with Indian accent, or any phonetic variant far from the dominant training data, the keyterm hallucination rate increases notably.

It's a classic distribution bias failure: STT models are trained predominantly on standard English audio, and when the input deviates from that distribution, the model "completes" with what it expects to hear rather than what it actually heard.

What the Filter Actually Does

Aditu doesn't replace the STT engine or modify the base model. It acts as a postprocessing step: it receives the transcription with potentially hallucinated keyterms and applies its filtering logic to decide which ones to keep and which to remove.

According to the author, in their internal tests the system reduces keyterm hallucinations by around 60%. It's not perfect, but it's a significant reduction for real-world use cases where incorrect transcriptions generate noise in downstream systems—CRMs, call analytics systems, semantic search pipelines.

Currently the service supports Hindi and English with Indian accent. It's available for free during the beta phase, and the author explicitly requests feedback on performance with your own data, as well as interest in features such as support for more languages, streaming, and self-hosting options.

Who Should Explore This

The clearest use case is any voice product deployed in non-English speaking markets or with multilingual user bases: contact centers in India or Latin America, voice assistants for sectors with specialized vocabulary, or automatic transcription systems for meetings with international participants.

It's also relevant for teams already using Claude with MCP servers connected to voice transcription APIs: a filter layer like this could be integrated as an intermediate tool in the pipeline, reducing noise before hallucinated keyterms contaminate the context that the language model receives.

The project is in very early stages—two points on HN and no comments at the time of publication—and the evaluation data is internal, which means you'd need to validate it with your own data before considering it for production. That said, the problem it addresses is real and documented, and the postprocessing filter approach makes sense as a pragmatic solution while STT providers themselves don't improve their robustness with non-standard accents.

From ClaudeWave we'll keep an eye on this: the combination of multilingual voice and agent pipelines is one of the most frequent friction points in the integrations we develop.