AI Statement

Last updated: 2026-11-02

1. In one paragraph

Rifflo uses machine learning in three different ways, and they should not be confused. The mastering engine is ordinary signal processing in your browser; a language model only reads your instruction and a few measurements and proposes a bounded change to its settings. The cloud tools run models on your file on a partner's servers to separate or clean it. The generator (text to audio) makes new sound that was not in your recording.

2. What the mastering assistant does, and what it does not do

It does:

• translate instructions such as "warmer", "less harsh" or "more punch" into changes of specific parameters;
• stay inside limits on each step (for example 2 dB on a tone control, 4 LU on the target, 0.3 on stereo width). Our server rejects an answer that goes past them;
• stay inside limits on the result (up to eight equaliser bands of 6 dB at most, compression ratio up to 3:1, stereo width from 0 to 150 percent, limiter ceiling from −3 to −0.1 dBTP, target from −24 to −6 LUFS). A result that would pass them is held at the limit;
• explain each change in a short card you can read, edit or undo.

It does not:

• hear or receive the audio. It works from measurements;
• generate sound, clone voices or imitate artists;
• make decisions about your account, your payments or your content.

3. Exactly what is sent to the model

When you send an instruction, our server forwards a request to OpenRouter, which routes it to the model provider. The request contains:

• your instruction text, up to 500 characters;
• up to six earlier instructions from the session, with the summary of what each one changed;
• measurements of your track, the current settings of the chain, the selected time range and the genre if you chose one;
• a few fixed example pairs that show the model the expected format.

It does not contain your audio, your file name, your email address or account identifier. Our server does not store the text: for signed-in users it keeps a log row with the model, status and token counts for 30 days. The text stays with your project in your browser, and it is included in a share link if you make one.

Other features send more to language models. "Suggest order" on a bio page sends its block types, links and labels.

4. Models currently in use

Versions are pinned in the code where the provider allows it. The echo remover, reverb remover and music noise removal use our own signal processing in your browser and no model. The licences page lists the licences of the audio models; two are a concern: BS-RoFormer states no licence, and the Demucs weights are released for research.

5. Training

We do not use your audio, stems, prompts or instructions to train, fine-tune or evaluate any model.

What we can say for our providers is narrower:

• Language-model requests to OpenRouter carry the setting "data_collection: deny", which OpenRouter documents as "use only providers which do not collect user data" (openrouter.ai/docs/features/provider-routing). OpenRouter's privacy policy states that OpenRouter itself does not use inputs or outputs for training (openrouter.ai/privacy).
• We run existing models on Replicate through its API and use no fine-tuning. Replicate's terms grant it a licence to process customer data only as needed to provide the service (replicate.com/terms).
• KIE's terms do not address this. We have no commitment from KIE. If that matters to you, do not use the lead and backing vocal splitter.

If we ever want to use user content to improve the product, we will ask you first with a separate opt-in, off by default. Changing this page or the Terms of Service alone is not enough.

6. The cloud tools and the generator

Separation estimates which parts of your mix belong to vocals, drums, bass and other instruments and writes each to its own file. Noise removal and enhancement estimate and change your recording; the enhancers can add detail that was not captured. Their output comes from your file, and none of them identifies who is singing or speaking or creates a voiceprint.

Text to audio makes new sound from your prompt. The result can resemble existing works, and it can be wrong. Check it before you use it. Files go where the Privacy Policy says, for the periods it states.

7. Labelling your work

A master made in the studio is your own recording with processing applied, and no part of the sound was generated by a model. In our understanding that is not "AI-generated content" under the rules we have reviewed:

• YouTube's disclosure rules state that using effects to enhance previously recorded audio does not need to be disclosed.
• The EU AI Act's marking duty for synthetic audio (Article 50(2)) does not apply to systems performing an assistive function for standard editing that does not substantially alter the input. We read mastering with Rifflo as falling within that exception; this is our reading, not legal advice.
• If a distributor asks, the accurate answer for a Rifflo master is: human recording and mix, AI-assisted selection of mastering settings, no generated audio.

For stems, denoised and enhanced files the picture is less settled: a model has re-estimated your recording. If a platform asks, say that an AI separation or enhancement model was applied to your own recording.

Samples from text to audio are AI-generated. Rifflo adds no marking to them, so disclosing that where the rules require it is up to you.

8. Known limitations

• The assistant works from measurements, not from listening. It can misread what you mean; if a change goes the wrong way, say so or undo it.
• It can only move the mastering chain within the limits in section 2. It cannot fix a mix or remove a specific sound, except that it can rebalance stems when you have agreed to separate them.
• Lossy files (MP3, AAC, Ogg) are decoded by your browser and decode slightly differently between browsers, so results for those files may differ marginally.
• Separation leaves artefacts, especially on reverb tails and overlapping instruments. Noise reduction can remove wanted detail. None of this is suitable for forensic or evidentiary use.
• Loudness targets are based on published platform guidelines, which change. Some are third-party measurements, and those are marked on the page.

9. Report a problem

If a tool produced an unsafe, offensive or clearly wrong result, write to support@voice-remover.com with the time of the request and, if you are comfortable sharing it, what you typed. We review every report within [5] business days.

10. Changes to this statement

We change this page when a model is added or removed, or when a commitment on it changes. A change that narrows a commitment about training takes effect only with the opt-in described in section 5.