RightVoice is Crele’s open benchmark for speaker isolation: it measures how well a tool keeps only the voice you asked for. For every 100 words your speaker says, a general speech enhancer still lets through 56 words from other people. Crele Focus 1.1 lets through fewer than 4.
What is RightVoice?
RightVoice is a public benchmark from Crele: give a system a two-person conversation and a few seconds of one voice, then score how much of that person comes back and how much of anyone else comes with them. Version 1 has 1,530 recordings, 16.45 hours, captured like phone and laptop audio.
Code on GitHub (Apache-2.0), data on Hugging Face (CC BY 4.0), results on the RightVoice page.
Why does keeping one voice need its own benchmark?
Because the usual failure is the wrong person, and speech benchmarks do not count it. Enhancement benchmarks score one voice in noise; separation benchmarks score waveform match. RightVoice counts words:
- Words lost. What your speaker said that is missing or wrong.
- Wrong-voice words. The other person’s words that come through.
- Talk over silence. Output while your speaker is quiet, including 121 recordings where they never speak.
What did RightVoice find?
A speech enhancer barely changes whose words you hear. DeepFilterNet3 lets through 56.2 wrong-voice words per 100 target words; the untouched recording lets through 61.7. USEF-TFGridNet, a research model that takes a voice sample, lets through 24.5. Crele Focus 1.1 lets through 3.7.
Where the target never speaks, the right output is silence. DeepFilterNet3 returns 69.5 words a minute there and USEF-TFGridNet 65.9, all of them from the other person. Crele Focus 1.1 returns 2.0.
| System | Wrong-voice words per 100 target words | Wrong-voice words per minute | Words per minute, target silent | Words per minute, target absent |
|---|---|---|---|---|
| Crele Focus 1.1Speaker isolation in the Crele app | 3.7 | 2.98 | 11.2 | 2.0 |
| USEF-TFGridNetResearch target-speaker model | 24.5 | 22.76 | 48.1 | 65.9 |
| DeepFilterNet3General speech enhancer | 56.2 | 47.35 | 85.8 | 69.5 |
Where does Crele Focus 1.1 not lead?
Crele Focus 1.1 is not clearly ahead on lost words: as released, 39.4% against USEF-TFGridNet’s 43.3%, under RightVoice’s 10% bar for “clearly better”. DeepFilterNet3 scores higher on voice match. Every measure is in the full table.
How do I test my own system on RightVoice?
Download the set, run your system on every item, then score the outputs.
Add --quick to score and report on a 192-clip subset first.
git clone https://github.com/Crele-AI/rightvoice && cd rightvoiceuv sync --extra scoreuv run hf download crele/rightvoice --repo-type dataset --local-dir data/build/v1# run your system: items/<id>/mixture.wav + sample.wav -> outputs/<id>.wavuv run rightvoice score --build data/build/v1 --system my_system --outputs outputsuv run rightvoice report --build data/build/v1Can I use RightVoice commercially?
Yes. The data is CC BY 4.0 and the code Apache-2.0. Cite it as “RightVoice (Crele)” with a link to the RightVoice page.
Questions
- Who made RightVoice?
- Crele. Crele's own speaker isolation model is in the results, so the data, the code and the scorer are public for anyone to check.
- Is there a RightVoice leaderboard?
- No, you are welcome to use the benchmark for your own evaluations.
- Which systems are in RightVoice v1?
- DeepFilterNet3, a general speech enhancer; USEF-TFGridNet, a research model that takes a voice sample; and Crele Focus 1.1, Crele's speaker isolation model.
- Does Crele train on RightVoice?
- No. Its readers, rooms and noise clips stay out of every Crele training set.
