Verified, then logged.
Sun 22 Nov 2026 · 00:20 KST / 15:20 UTC한국어 | ENSearch /

01 / AI · Analysis

Denoising audio before speaker embeddings raises error rates

The paper advises leaving input audio untouched and shrinking only the output vector, but its abstract gives no per-model error rates

A smartphone on speakerphone rests on a living-room coffee table while a television and a family member seen from behind blur in the background. AI-generated image

A benchmark of five open-source speaker embedding models under a sub-second latency budget found that denoising before embedding generally raised error rates for real-time voice agents. The paper appeared on arXiv, a preprint server, on 9 Nov 2026. The researchers also published a figure showing that roughly half of the TitaNet embedding is redundant for this task.

Real-time voice agents must separate the active user from speech in the background. According to the abstract, that speech passes through noise suppressors and causes false acceptances in speaker verification and speech recognition.

The paper's deployment guidance does not call for removing more noise. It calls for altering the audio sent to the speaker verification model less and shrinking only the output vector.

Background speech passes through the denoiser

The researchers compared TitaNet-Large, TitaNet-Small, ECAPA-TDNN, CAM++ and SpeakerNet. A speaker embedding model compresses the features of a voice into a numerical vector. The evaluation described in the abstract combined a controlled benchmark with a weighted, human-annotated production set.

The controlled benchmark was built from real calls, with a second real speaker injected into silent intervals at varying volume. The abstract names television audio, family members and colleagues as examples of background speech. Denoisers remove environmental noise effectively, it says, but background speech shares the spectral characteristics of the target voice and passes through.

The abstract calls denoising before embedding "generally harmful" and says it raised equal error rate across architectures. Equal error rate (EER) is the error rate at the point where false acceptances and false rejections are equal. The abstract gives no EER value for any individual model.

Guidance to keep raw, untrimmed audio

The audio-length test pointed the same way. According to the abstract, truncating audio below the full voice-activity segment degraded every model. The paper's guidance is to run speaker verification on raw audio in full voice-activity chunks.

The abstract's advice on decision thresholds reads: "Thresholds should favor precision over F1." Precision is the share of accepted decisions that are correct. F1 is a score that reflects both precision and recall.

Why half of the embedding is surplus

By the researchers' published figure, roughly half of the TitaNet embedding is redundant for this task. The abstract says the trailing dimensions encode population attributes such as language and gender. Those attributes carry little discriminative value in a user base with a similar demographic make-up.

Truncating audio degraded performance, according to the abstract, but truncating embeddings reduces the cost of similarity search. The abstract does not give the number of dimensions removed or the size of the saving.

The abstract does not say whether the same results apply to a demographically diverse user base. The benchmark's dividing line is where the reduction happens. The paper's guidance is to leave the input audio alone and shrink only the output vector.