VoxLock

VoxLock

Keeps the caller's voice and removes other voices from call audio before speech-to-text hears them. A 27.3M-parameter open model that streams on a CPU, with filters for Pipecat and LiveKit voice agents.

Made by arbazsiddiqui on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: VoxLock is an open voice isolation model for voice agents and phone calls. It takes one input, the call audio, and picks out the caller by itself. Other voices in the room are filtered out, so speech-to-text never receives them. It can run as a live stream on a CPU, with an algorithmic delay of 24 ms.

Who it suits: It suits developers who build voice agents or transcribe calls and need the caller's words to come through cleanly when other people are talking nearby.

Industry | Who uses it | How they use it | Example
Customer support | Teams running phone voice agents | Clean the call audio before transcription | Keep a caller's words when a coworker talks behind them
Voice AI | Developers using Pipecat or LiveKit | Add it as an audio filter or an Agents plugin | Put it in front of the speech-to-text step
Meetings and transcription | Teams transcribing headset audio | Isolate the headset wearer from other voices | Transcribe one person in a room full of talk

You can skip it if you need to keep several speakers at once, since it keeps one caller.

Where to use it:

• A Pipecat audio filter, for voice agents built on Pipecat
• A LiveKit Agents plugin, for agents built on LiveKit
• sherpa-onnx in C++
• The command line, on audio files, on Windows, macOS and Linux with Python

Use case | Who | What they get back
Voice agent front end | Agent developers | Caller audio with other voices removed
File cleanup | Analysts | A cleaned audio file of the caller
Transcription prep | Speech-to-text users | Audio that leaks fewer background words

How to set it up: You need these first:

• Python on Windows, macOS or Linux
• An audio file of a call, or a streaming audio source
• A voice agent framework, only if you want the Pipecat or LiveKit route

1. Install the voxlock package with pip.
2. Run the command line tool on a call recording.
3. For an agent, add the Pipecat filter or the LiveKit plugin, following the model card.

pip install voxlock

pip install -U huggingface_hub
hf download arbazsiddiqui/voxlock

How to use it: You run the voxlock command on a recording and name an output file, and the result holds only the caller. For your own streaming loop, or the agent plugins, the model card has more.

voxlock call.wav -o caller.wav

Tool | What it does
voxlock | Command line tool that isolates the caller's voice in an audio file
Pipecat filter | Audio filter that cleans call audio inside a Pipecat agent
LiveKit Agents plugin | Plugin that does the same inside a LiveKit agent
sherpa-onnx | C++ route for running the model

Example: You have a call recording where a television plays behind the caller. You run the command on call.wav and write caller.wav. The output is meant to keep the caller and drop the other voices.

For newcomers and experts: Voice isolation keeps one person's voice and removes the rest. A good first try is installing the package, running the command above on a short recording, and listening to both files.

For readers who know the field, the model has 27.3M parameters and starts from a causal REAL-TSE checkpoint. It is scored on how many background words a speech recognizer writes down while only the background is talking, counted per minute. The card's tests transcribe audio with Parakeet TDT 0.6B v2 and use the Krisp benchmark's own method.

Call audio comes in as one stream
VoxLock picks out the caller
Other voices are removed
Cleaned audio goes to speech-to-text
Transcript has fewer stray words

Good to know:

• Benchmark: the figures come from the author's run on 265 real recordings from the Krisp benchmark.
• Meetings: on AMI meetings, VoxLock had the lowest word error rate, but ai-coustics Voice Focus 2.2 S leaked fewer background words.
• Language: Hugging Face lists English.
• Size: 27.3M parameters, about 109 MB in F32.
• Caveat: the ElevenLabs result in the card covers only the 7 recordings with the most background speech.

Title: Word error rate on the Krisp benchmark, lower is better
VoxLock: 4.95%
Krisp VI 2.5 Default: 5.30%
ai-coustics Voice Focus 2.2 L: 6.84%
Sanas Voice Isolation SE 1.0: 9.38%
DeepFilterNet3: 13.30%
No processing: 23.21%

• License: Apache 2.0.
• Cost to run: none named. It runs on a CPU.
• Data it reaches: the audio you give it. The text available here names no network use.
• Not for: none named in the text available here.

On Hugging Face, 13 November 2026

• Task: audio to audio
• Parameters: 27.3M
• Weights: about 109 MB in F32; about 14 MB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 2,253
• Likes: 46
• License: apache-2.0
• Languages: en
• Based on: Wespeaker/wespeaker-voxceleb-ecapa-tdnn512
• Quantized versions: 1
• Tags: onnx, safetensors, voice-isolation, background-voice-cancellation, background-voice-removal, target-speaker-extraction, speaker-extraction, personalized-speech-enhancement, speech-enhancement, audio, speech, voice-agents

Links

• Model: https://huggingface.co/arbazsiddiqui/voxlock
• Files: https://huggingface.co/arbazsiddiqui/voxlock/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download arbazsiddiqui/voxlock

Then follow the usage in its model card. Nothing in this kit runs by itself.