decibel-hq/reso

This repository holds the training pipeline that Decibel Labs used to build Reso, a text-to-speech voice model. Text-to-speech means turning written text into spoken audio. Reso is described as expressive, streaming, and tag-controllable, meaning it can produce speech with emotional cues, generate audio as it goes rather than waiting for the whole sentence, and respond to inline tags placed in the text, such as laugh, chuckle, sigh, gasp, whisper, and pause, to change how a line is delivered. The README explains the pipeline in stages: a larger existing model, ElevenLabs, was used to generate many example audio clips from a written text corpus of over 16,000 phrases. Those clips were then checked automatically for quality, using a speech recognition tool to catch mistakes, and the clips that passed were used to fine-tune a smaller base model called Orpheus 3B through a technique called LoRA, which adjusts a small part of an existing model rather than retraining it from scratch. A status table in the README shows each of these stages has already been completed at least once, with rough numbers for how many clips were produced and passed quality checks. The repository is organized into folders for the text prompts, the numbered pipeline scripts, a database tracking pipeline state, data splits for training and testing, and saved checkpoints, though the actual audio files and trained model weights are deliberately left out of the repository. There is a quickstart for running a demo locally on an Apple M-series Mac, but it requires a separately obtained model checkpoint file that is not included in the repository. The project is released under the Apache 2.0 license, the same license as the base models it builds on.