Audio8 Realistic expands its speech generation to 23 languages
Edge0’s model accepts a reference voice and transcript, or generates speech without them.
Edge0’s model card lists 12 additional languages compared with Preview. To imitate a voice, the model needs a clean recording and an exact transcript of its words. The developer asks users to obtain permission for voice cloning. Language quality varies, and generated speech may omit or repeat words: listen before using it in a finished product.
Training for delivery
The research paper explains why recognizing the intended words and matching a speaker’s identity do not capture natural delivery. Training first learns from multilingual recordings, then rewards intelligibility and voice similarity, and finally adds a learned evaluator of realism. That evaluator compares generated speech with a human recording of the same sentence. Its scores rank alternatives for a particular prompt; they are not an absolute quality measure and do not replace checking the spoken content.
Deployment boundaries
For developers considering deployment, the runtime guide distinguishes Realistic from Preview: no Realistic ONNX export is available, and Preview’s exported weights are not interchangeable with it. The speech service also needs streaming enabled on the server before it can send audio in chunks during generation. With the default setting, it returns the complete audio after generation finishes, even if a request asks for streaming.
MarkTechPost also covered the release.
Sources(4)
- Release coverageChecked:
- Model cardChecked:
- ResearchChecked:
- DeploymentChecked:
