Resemblyzer is an open-source Python package from Resemble AI that analyzes and compares voices with deep learning. It has about 3,306 stars on GitHub under the Apache-2.0 license and was most recently active in 2023. Given a speech audio file, it produces a compact summary of the voice that machines can compare, cluster, and visualize.
At its core is a voice encoder: a deep learning model that turns an utterance into a 256-value embedding capturing the characteristics of the speaker’s voice. Two recordings of the same person produce similar embeddings, while different speakers land far apart, which makes many voice tasks straightforward. The repository ships runnable demos that show the main applications: speaker diarization, recognizing who is talking when using only a few seconds of reference audio per speaker; fake speech detection, comparing unknown utterances against reference audio to flag synthesized voices; and manifold visualization, projecting embeddings into 2D to see speakers cluster. Installation is a single pip command, and the package has minimal requirements, which is why it has been embedded in many research projects and voice tools. The underlying model was designed for speaker recognition rather than speech recognition: it ignores what was said and encodes how a voice sounds, including timbre and speaking style. That makes the embeddings useful for verifying identity, grouping recordings by speaker, and spotting synthetic speech in moderation pipelines.
Resemblyzer is free and open source under the Apache-2.0 license.
Resemblyzer is for speech researchers, audio ML engineers, and developers adding speaker recognition, diarization, or voice authenticity checks to their applications.