OmniVoice: open source text-to-speech and voice cloning for 600+ languages, 40 times faster than real time

The Next-gen Kaldi team has released OmniVoice, a text-to-speech model, under the Apache-2.0 licence. The project describes itself as a state-of-the-art massively multilingual zero-shot text-to-speech model and claims support for more than 600 languages. The GitHub repository has passed 10,700 stars and 1,700 forks.

Two capabilities: cloning and design

The model does two separate jobs. The first is voice cloning: it reproduces a voice from a sample you supply. The second is voice design: it produces a voice with the characteristics you ask for, without a sample. Zero-shot operation is part of the same picture, meaning no extra training is needed for a language or a voice the model has not heard before.

Speed

The repository gives a real-time factor of 0.025. One second of audio takes roughly 0.025 seconds to generate, which is 40 times faster than real time. With FlashInfer acceleration the project reports up to a 2.6x further speedup, reaching an RTF of 0.0115 at batch size 8. These are the project's own measurements; no independent evaluation has been published.

Who is behind it

k2-fsa runs Next-gen Kaldi, the successor to Kaldi, a long-standing name in speech processing. sherpa-onnx, from the same organisation, has 14,700 stars and runs speech-to-text, text-to-speech, speaker diarization and voice activity detection with no internet connection, supporting embedded systems, Android, iOS, HarmonyOS, Raspberry Pi and several NPU targets. What is on offer is therefore more than a single model: it is a stack already used to running on device.

For comparison

Inworld TTS-2, which we covered on 2 September, holds one voice identity across more than 200 languages and is offered as a commercial service. OmniVoice sits somewhere else: three times the language count, an Apache-2.0 licence, and downloadable weights, which means generation can stay on your own hardware. Language count and quality per language are different things; 600 is the number of supported languages, not a promise of equal quality in each.

Source: github.com/k2-fsa/OmniVoice.