AI voice converter

An AI voice converter changes the sound of the speaker in an existing recording using a reference voice. This is also called speech-to-speech voice conversion: your performance supplies the words, timing and expression, while the reference guides the new voice. It is useful for character dialogue, animation, game prototypes and alternate voiceover takes.

Record or choose a performance from 1 second to 5 minutes. Add a clean 5–10 second voice reference, or choose a voice from the library. The tool automatically extracts audio from videos and prepares compatible recordings on your device. Longer recordings ask you to select a section. Select Convert voice, listen to the preview and download the WAV.

Yes. Voice conversion is free and requires no signup, subscription or purchased credits. Processing uses your own device. The practical limits are recording length, available memory and processing speed; the first conversion also needs an internet connection to download the models.

Speech-to-speech starts with a spoken performance, so you control the words, pacing, pauses and expression by recording them. Text-to-speech starts with a written script and generates a performance. Use the voice converter when you already have a recording; use our Text to Speech tool when you want to start from text. Voice conversion does not translate or rewrite the script.

Yes. Choose a female or male reference voice independently of the speaker in your performance. You can filter the library by gender or use a reference recording you have permission to use. This uses AI voice conversion rather than a simple pitch shift. How closely the result matches the reference depends on the recordings.

Yes. Record a clean reference of your own voice, an original character performance, or another voice you have permission to use. Reuse the same reference across lines for a more consistent character. The tool uses that clip during conversion; it does not train a new custom model or create a saved voice profile. Listen to each result before using it in a finished project.

You can choose common audio files such as WAV, MP3, FLAC, M4A, AAC and OGG, or a video containing speech such as MP4, MOV, WebM or MKV. Support depends on the container, codec and browser. For recordings within the duration limit, the tool automatically extracts or prepares compatible audio locally. Longer recordings, or files whose duration cannot be read, ask you to select a bounded section first. Files need an audio track and must be no larger than about 600 MB.

The performance can be 1 second to 5 minutes long, and the voice reference must be 5–10 seconds. You can select a bounded section from a longer file. Performances over 20 seconds convert in overlapping sections with the same reference. Completed sections become available to preview and download while processing continues, followed by a full-length WAV. This is batch processing with partial previews.

Use 5–10 seconds of one clear speaker at a steady recording level. Avoid background music, echo, distortion, long silences and overlapping voices. A reference with a consistent vocal style gives the model a clearer target. For the performance, speak the words with the timing and expression you want in the result. Preview both recordings before converting.

The library has 15 English-speaking reference recordings with male and female filters. Names describe vocal texture, with options such as light and bright, warm and rounded, bold and textured, and dark and resonant. These are approximate descriptions guided by model estimates, rather than verified tone ratings or promises about the output. Preview the reference and demo to choose a voice; your performance supplies the delivery and emotion.

Accent transfer is unreliable. The model changes voice characteristics while pronunciation can remain close to your performance. The library is described by vocal tone rather than region. If you need a particular accent, record the performance with that pronunciation and preview the full conversion before using it.

The model uses your performance as the basis for the words, rhythm and expression, and longer results are assembled along the source timeline. It can still mispronounce, omit or alter words and change vocal details. Check names, emotional delivery and section transitions by listening to the full output. A matching automated transcript does not prove identical sound or expression.

English speech is the evaluated starting point, and the built-in references are English speakers. Other languages have not been validated for this tool. It does not translate speech into another language; if you need a different script or language, prepare that spoken performance first.

The current tool is intended for a single spoken voice. Singing, rapping, background music and overlapping speakers have not been validated. For a conversation, select one speaker’s section at a time. If the speech is noisy, clean it first and use a dry, clear reference rather than expecting voice conversion to remove music or repair a damaged recording.

The voice converter decodes, prepares and converts your selected recordings in your browser. It does not upload those recordings for inference. The site downloads the voice models and any public reference clips you choose. Cached models can be reused when browser storage allows; your own recordings are processed on your device.

You do not need to install an app or set up Python. The browser downloads about 496 MB of voice models on first conversion, plus the shared inference runtime. Downloads start when you convert, and verified models are cached when space allows. Clearing site data, storage eviction or blocked caching can require another download. Cached weights do not make this a guaranteed offline app.

A current desktop browser with plenty of free memory is the best starting point; the tool has been checked in desktop Chromium. It detects GPU availability automatically and falls back to CPU processing when needed. A phone or a low-memory computer may struggle with the download, model initialization or long recordings.

The model runs on your own computer, so speed depends on its processor, GPU support, available memory and recording length. CPU processing can take several minutes per section, and a short recording can take longer to convert than it takes to play. The first run also downloads and prepares the models. Keep the tab open, start with a short test and use the progress display to follow longer jobs.

This tool converts recorded speech and produces a downloadable audio file. It does not provide a virtual microphone or change your live microphone feed in Discord, Zoom or a streaming app. The previews for longer recordings are completed batches, not a real-time audio stream. You can use the finished WAV in an editor or another app that accepts audio files.

Record requests the microphone selected by your browser and operating system. A connected phone or external microphone can be chosen as that input. Select your preferred microphone in the browser or system sound settings, then stop and restart recording. The voice converter does not explicitly request a nearby phone and currently has no microphone picker.

The output is a mono, 24 kHz, 16-bit PCM WAV file. You can listen before downloading and use it in an audio editor, video editor or game project. The tool exports the converted audio, including when the input was a video; it does not replace the soundtrack inside the original video or currently offer an MP3 export.

The library uses CSTR VCTK (CC BY 4.0) and LibriVox recordings (public domain in the US). Brief credits and source links appear in the voice library. Use your own recording or a reference you have permission to use. Recording reuse rights do not automatically establish unrestricted permission to impersonate a speaker, imply endorsement or use their identity. Check the source terms for your intended project.