Zyphra Zonos

Zonos TTS copies a voice from a reference clip of roughly 10 to 30 seconds, with no extra training. Give it clean audio and a text prompt, and it predicts how that speaker would read your script. Zyphra reports that reviewers found short-sample clones convincing on first listen, but pacing can drift on longer passages.

Beyond speaker embeddings, you can feed an audio prefix alongside your text. This nudges the model toward harder-to-copy behaviors such as whispering, which embeddings alone struggle to reproduce. It's a practical tool when the exact delivery matters more than raw similarity.

The conditioning setup exposes speaking rate, pitch variation, maximum frequency, and audio quality as adjustable knobs. Emotions including happiness, anger, sadness, and fear are tunable too. For anyone producing character dialogue or dramatic reads, that control is the reason to pick Zonos over a fixed-voice generator.

Zonos v0.1 handles English, Japanese, Chinese, French, and German. English dominates the 200,000-plus hours of training data, so expect the crispest results there, with the other languages solid but less polished. It's enough for multilingual projects that would otherwise need several separate tools.

You choose between a Transformer backbone and a Transformer-Mamba hybrid. Zyphra says the hybrid was the first TTS model to combine these designs. Both use the same weights-per-parameter scale, so the pick comes down to how each behaves on your hardware and content.

The Apache 2.0 weights run fully offline, and a packaged Docker file handles setup. Zyphra says the model hits a real-time factor of about 2x on an RTX 4090, meaning it generates two seconds of audio per second of compute. Running locally keeps audio private and removes ongoing per-use charges.