AudioShake Debuts Multi-Speaker 2.0 for Separating Overlapping Voices

AudioShake has released Multi-Speaker 2.0, the company's newest Multi-Speaker model, which removes 75% more noise and offers 32% better separation of overlapping voices than its previous model. In what the company is calling the world's first low- and high-resolution multi-speaker voice technology, Multi-Speaker 2.0 pairs high-accuracy diarization with benchmarked sound separation, turning conversational audio and overlapping voices into clean, isolated tracks.

Multi-Speaker Separation takes a fully mixed track and pulls each person onto their own isolated, clean track. From podcast interviews and reality TV to phone calls, interviews, and field recordings, much of the world's audio is filled with conversation and overlapping speakers.

Multi-Speaker 2.0 also delivers 4.1 fewer transcription errors than the best open-source separator, 30% lower diarization error than pyannote community-1, and lower error than the paid service it replaced in every domain the company tested: meetings, telephone, broadcast, and web video. Additionally, the separated audio comes back scored per frame and per file, so users can process high-confidence material automatically. When a conversation is sent straight to a transcription tool, errors climb as people interrupt one another. Running Multi-Speaker 2.0 first flattens that climb, cutting errors in half at the heaviest overlap.

Built for Real-World Audio

Multi-Speaker 2.0 uses a single model across sample rates from 8 kHz to 48 kHz, covering phone calls and meeting captures through podcast, radio, and film/TV masters, with no pre-processing or format conversion required. It is designed to handle far-field recordings with room reverb and crosstalk, as well as location and archival dialogue buried under traffic, music, and crowd noise. On telephone audio, diarization error is 38% lower than pyannote community-1.

Because it is an acoustic model rather than a language model, Multi-Speaker 2.0 works on how voices sound, not what is being said. It is trained across a wide range of languages, including accented and code-switched speech. Accuracy holds up to roughly 20% sustained overlap (a fifth of total speech time with two or more people talking at once) before it begins to degrade. At that level of crosstalk, according to AudioShake, separated tracks transcribe as accurately as individually mic'd speakers.

Diarization and Confidence Scoring

Alongside the separated tracks, each job returns a structured results file with an overlap-aware timeline of who spoke when. When two people talk simultaneously, both are marked active rather than one being dropped, preserving interjections and crosstalk for captioning, subtitling, dubbing, and data prep.

Every job also returns two confidence scores: one measuring whether speech was assigned to the correct speaker and one measuring how cleanly overlapping voices were separated. Each is scored from 0 to 1 per 20-ms frame and as a whole-file score weighted by speech activity. Frame-level scores point QC reviewers to specific timestamps, while file-level scores rank a batch so that high-confidence files move through the pipeline untouched and only uncertain ones are flagged for review.

Applications and Availability

AudioShake is aiming Multi-Speaker 2.0 at film/TV post-production (isolating overlapping dialogue for editing, ADR, and remixing), podcasting, dubbing and localization, transcription, voice-AI training data, and accessibility workflows such as captioning.

The technology is available through AudioShake Live, the company's drag-and-drop web platform for media and production teams, and through the AudioShake API with diarization and confidence scores included. It is not currently offered through the AudioShake SDK, which is reserved for real-time, on-device products.

LINK:	https://www.sportsvideo.org/2026/09/24/audioshake-debuts-multi-speaker...
	See more stories from svg

Most recent headlines