Inflect Mini v3: A small English text-to-speech model that runs on CPU

English text-to-speech v3 uploaded by Owen Song on 11 10, the model card states the core has 34.31M parameters (138.4 MB with FP32 weights), native 24 kHz output, and an optional enhancer with 2.85M parameters that outputs 48 kHz. It includes six built-in voices: cassidy, mark, hope, james, sophia, and jasper, supporting 8 types of voice event tags such as [laugh], [sigh], and [gasp]; Voice Tuner can create a reusable voice profile from authorized recordings of 5–15 seconds. The repository is a complete inference package under the Apache-2.0 license.

Original title: owensong/Inflect-Mini-v3

English text-to-speech v3 uploaded by Owen Song on 11 10, the model card states the core has 34.31M parameters (138.4 MB with FP32 weights), native 24 kHz output, and an optional enhancer with 2.85M parameters that outputs 48 kHz. It includes six built-in voices: cassidy, mark, hope, james, sophia, and jasper, supporting 8 types of voice event tags such as [laugh], [sigh], and [gasp]; Voice Tuner can create a reusable voice profile from authorized recordings of 5–15 seconds. The repository is a complete inference package under the Apache-2.0 license.

Why it matters Small size, Apache license allows commercial use, and it can generate locally on CPU, making it suitable for narration, operational prompts, and offline reading

Rising score 42 · Likes 42 · Downloads 590

Reasons

  • Author's own evaluation: average word error rate of 0.739% across six voices (1,688 text segments per voice), UTMOS22Strong 4.479; the author notes this is not an official leaderboard submission and cannot be directly compared with leaderboard numbers from Kokoro, etc.
  • The author claims approximately 162x real-time speed on RTX 5090 using an optimized CUDA path, and approximately 1.81x real-time speed on Ryzen 9 3900X using 2 threads
  • The same series also includes Micro v3 with 11.79M parameters and Atom v3 with 157.19K parameters; GitHub also provides a training toolkit to adapt Mini to other languages

Usage

  • Python 3.12 new environment: use hf download owensong/Inflect-Mini-v3 to download the entire repository, run verify_release.py for verification, install core/requirements-cpu.txt, then run python run.py speech --voice cassidy --text "..." --output out.wav
  • Voice events: first install dependencies from extras/events, then run python run.py events --voice cassidy --text "... [laugh] ..." (only available on CPU at 24 kHz); use --profile enhanced for 48 kHz
  • GPU: create a separate environment and install core/requirements-cuda.txt, add --device cuda --runtime cuda-source; online demo is available in the owensong/Inflect-v3 Space

Risks

  • Supports English only; event tags do not change the emotion of surrounding sentences, and there is no style control for whisper, anger, etc.
  • Voice Tuner only supports registration on Linux CPU, similarity depends on the recording, and the author states that replicating the original voice is not guaranteed; only authorized voices can be used
  • Word error rate and audio quality scores are from the author's own evaluation, and the evaluation corpus was used during development; the enhancer may produce slight hum on some materials. This entry has not been run

Open

Hugging Face Trending

Explore the latest AI Radar updates