Datasets:

Krisp-AI
/

VoiceIsolation-Benchmark-Dataset-Processed

configs:
  - config_name: work
    data_dir: work
  - config_name: call_center
    data_dir: call_center
  - config_name: phone_calls
    data_dir: phone_calls
license: cc-by-nc-4.0
task_categories:
  - audio-classification
tags:
  - voice-isolation
  - speech
  - noise-reduction
pretty_name: Voice Isolation Benchmark - Processed Audios

Voice Isolation Benchmark – Processed Audios

Audio samples processed by four Krisp Voice Isolation models. Each scenario folder contains subfolders for every model, with one processed .wav file per original sample.

Voice Isolation Models

VI 2.5 Default (`vi_2_5_default`)

The main Voice Isolation model. Strongest at removing noise and other speakers. Goes fully silent when only a bystander is talking.

VI 2.5 Balanced (`vi_2_5_balanced`)

Strongly removes noise and background voices while applying lighter processing when the main speaker is talking. This keeps their voice sounding natural, but a small amount of background sound may still come through.

VI 2.5 Lite (`vi_2_5_lite`)

Part of the 2.5 family, requiring about a third of the compute. While the full VI 2.5 Default remains the recommended choice for the hardest cases, Lite provides comparable quality on average, cutting errors by ~68% on the full benchmark.

VI 2.5 HD (`vi_2_5_hd`)

Similar to VI 2.5 Default, but designed for 32 kHz high-fidelity audio streams (such as WebRTC or WebSocket applications) rather than narrowband telephony codecs.

Audio Format

All processed files are uncompressed WAV (PCM 16-bit signed little-endian), mono channel. Default, Balanced, and Lite models operate on 16 kHz sampling rate. HD processes the full 32 kHz stream, preserving the original sample rate.

What is released

This dataset contains the output of four Krisp Voice Isolation models applied to every recording in the Voice Isolation Benchmark Dataset. Benchmark results across 11 STT configurations are published as a Space, where the evaluation methodology (corpus-level WER via jiwer, DNSMOS-C NISQA on primary segments) is documented.

Folder Structure

Processed_Audios/
├── call_center/
│   ├── metadata.jsonl
│   ├── vi_2_5_default/
│   ├── vi_2_5_balanced/
│   ├── vi_2_5_lite/
│   └── vi_2_5_hd/
├── phone_calls/
│   ├── metadata.jsonl
│   ├── vi_2_5_default/
│   ├── vi_2_5_balanced/
│   ├── vi_2_5_lite/
│   └── vi_2_5_hd/
└── work/
    ├── metadata.jsonl
    ├── vi_2_5_default/
    ├── vi_2_5_balanced/
    ├── vi_2_5_lite/
    └── vi_2_5_hd/

Voice Isolation Benchmark – Processed Audios

Voice Isolation Models

VI 2.5 Default (vi_2_5_default)

VI 2.5 Balanced (vi_2_5_balanced)

VI 2.5 Lite (vi_2_5_lite)

VI 2.5 HD (vi_2_5_hd)