VoxLock extracts a target speaker for voice agents without enrollment
Original title: VoxLock: Target-Speaker Extraction for Voice Agents Without an Enrollment Step
On Krisp's 265-recording set it reaches 4.95% WER versus 23.21% unprocessed, with 24 ms latency.
A fine-tune of the public causal REAL-TSE model, 27 million parameters, uses a voice sample cut from the call itself so a recognizer does not transcribe nearby speakers. On Krisp's 265 recordings the author reports 4.95% word error rate against 23.21% unprocessed, and 9 background words over 85 minutes of background-only speech. On 400 frozen AMI meeting windows, word error rate falls from 42.95% to 10.63%. The 8-bit model uses 77% of one CPU thread with 24 ms algorithmic latency; model, code and per-file scores are open.
Unconfirmed: Word-error comparisons and significance claims are the author's own measurements.
- REAL-TSE was fine-tuned on one GPU and released as VoxLock.
- Author reports WER not significantly different from Krisp's best filters.
- WER is significantly lower than the other commercial and open systems tested.
- The 8-bit model runs a live call with 24 ms algorithmic latency.
- Organizations
- Krisp
From the original article
1 min read at huggingface.co
A voice agent's speech recognizer transcribes everyone near the microphone, including people other than the caller. I fine-tuned a public causal target-speaker extraction model (REAL-TSE, 27M parameters) on one GPU to work from a voice sample cut from the call itself, and released it as VoxLock. I score it by what a recognizer writes on real recordings.
By Arbaz Siddiqui · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.