Which AI voice model should you use?
Start with what you want to make. Voice conversion models transform a performance you record, singing models follow an existing vocal, and text-to-speech models generate speech from a script. The right framework depends on that job and the software you plan to use.
Start with the audio you want to create
Change a live or recorded performance
Choose a voice-conversion model when you want to keep the words, timing, and expression of an existing performance while changing its vocal character.
Convert an existing vocal track
Choose a singing voice-conversion model when melody, pitch, and the original singer's phrasing need to drive the result.
Turn a written script into speech
Choose a TTS model when you want to type dialogue, narration, or another script and generate spoken audio without first recording it.
For voice changing: start with RVC or Beatrice
RVC is the largest voice-conversion ecosystem currently indexed by AIVoices. It is used for recorded conversion and in compatible real-time applications. A typical model uses a PTH weight and may include a retrieval index. The model still needs an input voice; it does not generate a performance from text by itself.
Beatrice is another real-time voice-conversion framework supported by compatible versions of VCClient. Beatrice models are a separate format, so an RVC model cannot be relabeled and loaded as Beatrice. AIVoices will expose its directory once verified public inventory is available.
For vocal covers: consider so-vits-svc
so-vits-svc is built for singing voice conversion rather than conventional text-to-speech. It converts a source vocal and follows its melody and delivery. An unpacked model commonly needs a generator checkpoint and matching configuration, while some creators package the required files together in a ZIP.
Framework versions and configurations matter. Use the instructions on the original model page instead of assuming that every checkpoint works in every so-vits-svc application.
For text-to-speech: choose GPT-SoVITS or Piper
GPT-SoVITS supports speech generation and voice-cloning workflows that use reference audio. Model packages commonly contain separate GPT and SoVITS checkpoints, so keep the documented pair together and follow the matching software-version instructions.
Piper is a fast local text-to-speech engine. Piper voices normally use an ONNX model plus a matching JSON configuration and are released for specific languages and quality levels. It is a practical fit for offline narration, assistants, and accessibility projects when a suitable voice and license are available.
Check these details before downloading any model
- Intended use: voice conversion, singing conversion, and TTS are different jobs.
- Framework and version: confirm that your chosen application supports the exact model format.
- Complete files: keep paired checkpoints, configs, and indexes together when the framework requires them.
- Language and samples: use documented language support and demos rather than guessing from the voice name.
- License and source: review the uploader's current terms before publishing or monetizing output.
Sources and further reading
- RVC Project repository
- so-vits-svc 4.1 documentation
- GPT-SoVITS repository
- Piper repository
- VCClient framework support documentation
Reviewed against the linked project documentation on July 20, 2026. Update this guide when supported formats, file layouts, or AIVoices framework coverage materially changes.