llama.cpp added support for Soprano speech models in build b11850. The integration converts the model’s text-processing backbone and its audio decoder into the formats used by the local inference software. It also adds preprocessing and memory accounting. The decoder processes the whole utterance, so streaming mode delivers the audio in a single chunk after the utterance is complete.
llama.cpp, software for running models locally, added Soprano speech generation in build b11850. The project’s release notes list Soprano-80M and Soprano-1.1-80M support. The integration combines a Qwen3 backbone, which handles the model computation, with a Vocos decoder that produces audio. It includes the conversion work needed to export the backbone and decoder weights into the software’s model formats.[1]
The Vocos decoder processes the complete utterance. In streaming mode, audio therefore arrives as one chunk after the flush operation, rather than as progressively decoded pieces. The decoder loads as a generation-only mmproj component, separate from the exported Qwen3 backbone. The helper pipeline connects these parts to the text-to-speech tool, whose documentation now includes Soprano usage.[1]
Before generation, the integration prepares Soprano text by making letters lowercase, combining whitespace and handling each digit separately. Unknown input falls back to an unknown token. Memory accounting now includes the context needed for audio generation. The audio implementation adds Vocos operations together with the inverse Fourier transformation over short time windows used in decoding. The project’s own release notes supply these implementation details.[1]
