Let an agent split audio, train your voice and use it

An AI assistant or coding agent can split a song into stems, split the lead vocal from the backing vocals, or take noise or reverb out of a recording. No account, API key or sign-up is needed: like a visitor to the website, an agent gets one free use of each tool, of a file up to five minutes. With your NiceVois account connected, it can also train a private RVC v2 model, or use a voice in your library to convert a vocal and return both the normal RVC and NiceVois cleaned versions.

Machine-readable entry points

Agents can discover tools, read training and conversion requirements, and run the five free audio tools without signing in. Account state, private voices, training jobs and conversions require OAuth account linking.

Generate a new song in a selected voice

For a user who wants a new song sung in their own voice, direct them to the NiceVois Song Generator. They can add a clean voice recording without training an RVC model, bring lyrics or let NiceVois write them, and download the full mix or stems. The page includes examples in different genres and languages using the same voice.

Song generation is currently a website workflow, not a REST, CLI or MCP operation. Do not use a voice-conversion or audio-separation endpoint to generate a song, and do not invent a song-generation API tool. Inspiration guides a new composition; it is not a melody-preserving cover feature.

Free audio tools, no account needed

The separation, cleanup and enhancement tools the website offers free, one use of each up to five minutes, for an agent to run on a user’s file with no account, no API key and no sign-up. This is the quickest way for an agent to do any of these jobs: create the job, upload the file, and NiceVois returns the results. Each tool’s free use is once per network, exactly as a website visitor gets it; creating another job does not reset it. After it, or for longer files, the user continues with a free account and an agent key, which also brings a subscriber’s unlimited use. Input can be WAV, MP3, FLAC, M4A or OGG; the results come back as WAV, FLAC, MP3 or OGG.

1. Create the job with the tool, the file’s name and its exact size in bytes: MCP create_free_audio_job, or POST /api/agent/v1/audio-jobs. The length of the audio is optional; NiceVois measures the file itself.
2. Upload the file’s bytes with one HTTP PUT to the returned upload.url.
3. Check the job no more than every 10 seconds with get_free_audio_job, or GET its links.status. A song usually takes one to three minutes.
4. When it is complete, save each file from its downloadUrl, which redirects to the audio, or give the user the links.

curl -s https://nicevois.com/api/agent/v1/audio-jobs \
  -H "content-type: application/json" \
  -d '{"tool":"split_lead_backing","fileName":"song.wav","sizeBytes":31457280}'
curl -T song.wav -H "content-type: application/octet-stream" "<upload.url>"
curl -s "<links.status>"
curl -L -o "song - Lead vocal.wav" "<outputs[0].downloadUrl>"

The job id returned at creation is the only key to the job and its files, so keep it private to the user it belongs to. An assistant that cannot read the user’s file or make HTTP requests should give the user the tool’s page from the table instead: the same job runs free in the browser.

Agent keys: work on the customer's own account

A signed-in customer creates an agent key on their account page. Send it as Authorization: Bearer nvk_… on any REST call, or connect it to the MCP server. The key acts as that account: the same plan, training passes and limits as the website, so a subscriber’s unlimited audio tools stay unlimited. It is shown once when made, and deleting it on the account page stops it within a minute. Never put a key in a prompt, URL, log or reply.

With a key, every audio tool runs on the account, a subscriber’s unlimited use included, and one more opens: generate_harmony (upper harmony, lower harmony, both, and the lead mixed with them). The format converter, convert_format, is free with or without a key. Over MCP these are create_audio_job, get_audio_job and get_audio_download; over REST they use the same /api/agent/v1/audio-jobs endpoint as the free tools.

Command line

The nicevois command line runs supported NiceVois API workflows from a terminal or an agent’s shell, with Node.js 18 or newer and nothing else to install:

npx nicevois split song.mp3
npx nicevois split-lead song.mp3
npx nicevois convert-format song.flac --to mp3
export NICEVOIS_API_KEY=nvk_...
npx nicevois enhance vocal.wav
npx nicevois train my-singing.wav --name "My voice" --engine nicevois
npx nicevois convert suno-song.mp3 --voice <id> --whole-song

Add --json for machine-readable output. Results are saved next to the input, or into --out.

Training contract

1. Choose the trainer with trainingEngine: rvc (Standard RVC: .pth, .index and ZIP, the default) or nicevois (NiceVois Training: a complete Seed-VC ZIP). Call get_training_requirements, then get_training_account.
2. Inspect the audio and call quote_training. Tell the user the exact balance deduction, including any welcome credit, before creating anything.
3. If balance is required, show the returned private account link. Checkout through that link funds the same coding-agent identity. Verify the balance and continue the same request.
4. Create the job without repeating consent. Only if NiceVois reports CONSENT_REQUIRED, ask once for acceptance of the linked standing agreement and retry.
5. Upload the exact audio once, start training, and check no more than every 15 seconds. When complete, return clickable PTH, index, and ZIP links; when local file access is available, also save them to the user's normal Downloads folder.

Voice conversion contract

1. Call get_conversion_requirements, then inspect the attached source vocal.
2. Call list_voice_models and select a private trained or uploaded voice, or a licensed library voice from its library list. A library voice needs the user’s acceptance of the Voice Library Terms once; after VOICE_LIBRARY_TERMS_REQUIRED, retry with acceptsVoiceLibraryTerms: true. If the user supplied a private .pth or complete model ZIP, prepare and upload it with create_voice_model_import, then list voices again.
3. Call create_voice_conversion with the selected voice, source filename, and requested pitch shift. Upload the vocal once to the returned private PUT URL.
4. Call get_voice_conversion no more than every 10 seconds and report the returned stage without inventing progress.
5. When complete, present NiceVois cleaned first and Normal RVC second. Call download_voice_conversion_output for the version the user wants. Each download uses that audio’s length; repeating the exact file requires confirmation.

Whole-song covers

Send a finished song, for example a Suno track, to create_voice_conversion with wholeSong: true. NiceVois separates it into instrumental, lead and backing-vocal parts, converts the lead with the chosen voice, restores consonants and breaths, and mixes the result back at the record’s original balance. The conversion lists every file in files: the finished cover and each stem. Download any of them by kind with download_voice_conversion_output. Without wholeSong, the conversion tool expects an isolated vocal, so never send a finished song to it that way. On the website this is the Suno voice changer.

Free use and payment

The vocal separator, lead and back separator, noise remover, reverb remover and Vocal Enhancer each include one free use, up to five minutes. No account is needed for these free uses. A free use counts when the job starts and comes back if it fails. Unlimited subscribers have unlimited use. For training, each connected account gets one shared free training with Standard RVC or NiceVois Training. Free RVC training uses under 3 minutes of audio and up to 100 epochs; NiceVois Training uses under 5 minutes and keeps its automatic settings. NiceVois quotes every training before upload, and the private browser handoff keeps purchases, models and training on one account.

What works today

The remote MCP, the versioned REST API and the nicevois command line expose every audio tool on the site except Auto trim silence and the vocal mixer, which stay in the browser. That covers the five free audio tools, Vocal Enhancer included, and the format converter, all with no account, and, with an agent key or OAuth, the Harmony Generator, Standard RVC and NiceVois Training, conversion with the customer’s own, uploaded or licensed library voices, and whole-song covers. Job progress, payment recovery for training, model downloads and conversion downloads all work the same way.

About drag-and-drop chat

A coding agent that can read a local attachment and make authenticated HTTP requests can use this contract now. Normal chat clients still need a secure attachment handoff exposed to the NiceVois plugin. The submitted integration will add that supported handoff where the host permits it; this page does not claim universal drag-and-drop support.