Inflect-Mini-v3 / core /README.md
owensong's picture
Publish official Apache-2.0 model card, feature guides and notices
9f539a2 verified
|
Raw History Blame Contribute Delete
1.96 kB

Inflect Mini v3 Core

Local English speech with six preset voices and native 24 kHz output. Use the model card for setup, voice samples and public commands.

Speech and Audio

The synthesis core has 34,310,790 parameters. Its six voice IDs are hope, sophia, cassidy, james, jasper and mark. Native 24 kHz speech is the default. A separate compact enhancement component provides optional 48 kHz speech; it does not change the backbone's native rate. Frontend resources, voice-conditioning assets and optional components are separate from the core parameter count.

Low-Level Interface

From this directory, with the Python 3.12 requirements installed:

python synthesize.py --model mini --voice cassidy --text "Welcome back." --device cpu --output ../../mini.wav
python synthesize.py --model mini --voice hope --profile enhanced --text "The next chapter starts here." --output ../../mini-48k.wav

The public run.py in the repository root is the recommended entrypoint. Choose a new WAV outside the downloaded package. CPU and the canonical runtime are defaults; CUDA requires the CUDA requirements and --device cuda. Use --subtype PCM_16 for PCM16 instead of floating-point WAV. The CLI reports the applied text handling and processing configuration.

Event tags and VoiceTuner are supplied separately in ../extras/ and documented on the model card. They are not emotion controls embedded in ordinary speech.

License

Inflect-owned source and the weights identified in LICENSE_SCOPE.json are licensed under Apache-2.0. The repository's licensing guide explains scope and retained historical notices. Third-party components retain their own licenses; see notices and LICENSES/.

assembly-manifest.json records this Core's integrity inventory. Historical source information under provenance/ is preserved separately.