VoiceFilter
Google's target-speaker extractor: give it a short reference recording and it isolates that one voice from a noisy multi-speaker mix.
About
VoiceFilter solves a different problem from plain denoising: given a cocktail-party mix, pull out one specific person's voice. It trains two networks separately — an LSTM speaker encoder that produces a discriminative speaker embedding from a reference recording, and a masking network that takes the noisy spectrogram plus that embedding and predicts a mask for the target speaker's voice. Speaker-conditioning sidesteps the three classic headaches of blind separation: unknown speaker count, the permutation problem, and deciding which output belongs to whom. The original 2018/2019 paper (Interspeech 2019) framed it as a front-end for speech recognition and reported large WER reductions on multi-speaker audio with minimal damage to single-speaker input. Google followed with VoiceFilter-Lite (on-device, streaming) and the approach is covered by Google patents. There is no official public release of Google's model — community reimplementations exist — so this is influence and design more than downloadable artifact. The idea, however, is everywhere now: every 'focus on this speaker' feature in conferencing products is a descendant.
Specs
Context window
Pricing
Free quota
API access
Weights
Released
Benchmarks
- Paper (Interspeech 2019, author-reported): large WER reductions on multi-speaker test sets with minimal single-speaker degradation
- No independent benchmark replication found; no public model to benchmark
Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.
Use cases
Trust score — 7/10
Scored 7/10 (TIER metrics, metrics-v1). Evidence: Google-authored Interspeech 2019 paper with a clean, influential formulation of speaker-conditioned extraction; followed by VoiceFilter-Lite and patent filings; one of the most cited designs in target-speaker work. Strengths: solves the permutation and speaker-selection problems elegantly, real downstream use inside Google's speech stack (company-reported). Docked for: (1) no public code or weights from Google — nothing to inspect or reproduce; (2) all evaluation numbers are author-reported. License: none — no public release.
How the trust score is built
Full methodology →- Adoption · stars/downloads, log-scaled vs peers
- 9.0/10
- Maintenance · recency, releases, issue load
- 5.0/10
- Openness · license + docs
- 3.0/10
- Manual review · sanity read + red-flag check
- 7.0/10
- · Google; arXiv 1810.04826 (Oct 2018); Interspeech 2019
- · Target-speaker extraction via speaker-conditioned masking: LSTM speaker encoder + masking network
- · Large WER reductions on multi-speaker input (author-reported); followed by VoiceFilter-Lite + patents; no public weights