Release of Inflect v3 Text-to-Speech Model Family

Developer Owen Wang has announced the release of Inflect v3, a new family of lightweight, local text-to-speech (TTS) models. The release includes three distinct model sizes: Mini, Micro, and Atom, designed to operate on local hardware via CPU or CUDA.

Model Specifications and Capabilities

The primary model in the release, Inflect Mini, contains 34.31 million parameters and includes six distinct voices. According to documentation provided on GitHub, the Mini model supports 48 kHz audio enhancement and features “reference-voice tuning,” a capability that allows users to generate speech based on a provided audio profile, though the developer notes it may not replicate every speaker with total accuracy.

Inflect Mini also introduces support for specific vocal-event tags, allowing users to insert non-verbal sounds such as [laugh], [sigh], [gasp], [cough], [yawn], [inhale], [exhale], and [throat_clear]. The developer noted that these tags currently utilize repeatable performances, which may result in limited variation when tags are used repeatedly.

The broader model family includes:

  • Mini: 34.31M parameters, focused on audio quality and feature set.
  • Micro: 11.79M parameters, a smaller neural speech model.
  • Atom: 157K parameters, described as an extremely small learned core.

The weights and inference code for the models have been released under the Apache 2.0 license.

Early User Feedback and Technical Observations

Following the announcement on Reddit, early users have provided preliminary assessments of the model’s performance. Some users have suggested that the quality of the Mini model exceeds that of other existing local TTS options, such as KokoroTTS.

However, some technical observations regarding audio fidelity have been noted by community members. One user reported the presence of high-pitched static or artifacts in the generated audio. The user suggested that these artifacts appear more noticeable in lower-toned male voices, such as “Cassidy,” “Mark,” “James,” and “Jasper,” while being more imperceptible in higher-pitched female voices like “Hope” and “Sophia.” These observations remain unverified by independent technical benchmarks.

Availability

The Inflect v3 model collection is hosted on Hugging Face, with the source code and installation instructions available via GitHub. The developer has requested that users report any errors in sentence synthesis, including the specific text, voice, and settings used, to assist in further refinement.