Inflect-Micro-v2 ships standalone 9.36M parameter TTS

// 1h agoMODEL RELEASE

Inflect-Micro-v2 ships standalone 9.36M parameter TTS

Inflect-Micro-v2, released by developer Owen Song on Hugging Face, is an open-weight text-to-speech (TTS) model designed for efficient local execution. Packing a complete end-to-end speech generation pipeline into 9.36 million parameters (~37.5 MB FP32), it generates 24 kHz English speech directly on CPU or CUDA hardware without external vocoders.

// ANALYSIS

Shrinking complete end-to-end speech generation models under 10 million parameters unlocks real-time, privacy-focused voice synthesis on low-power edge hardware.

–High Efficiency: At just 9.36M parameters, the model enables rapid local inference on standard CPUs and GPUs with minimal memory consumption.
–Standalone Architecture: Integrates text processing, timing prediction, and waveform generation without requiring external vocoder models.
–Open & Accessible: Published under an open license on Hugging Face with an interactive space for immediate browser testing.

// TAGS

speechttsopen-sourceaudioedge-aihuggingfacellm

DISCOVERED

1h ago

2026-07-26

PUBLISHED

5h ago

2026-07-26

RELEVANCE

8/ 10

AUTHOR

nateb2022

// KEEP READING

More AI developer news from the feed

EXPLORE FULL FEED

MODEL1h ago

Black Forest Labs previews multimodal model Flux 3

Black Forest Labs has previewed Flux 3, a unified multimodal foundation model designed to natively integrate image creation, audio synthesis, 720p video generation with up to 20 seconds of synchronized sound, and robotics action prediction. Early access features text-to-video, image-to-video, and keyframe transitions, with an open-weight community release planned.

OPEN SOURCE1h ago

Homie brings multi-view consistency to AI video

Homie is an open-source reference-to-video framework designed to solve subject and object identity drift in AI video generation. By leveraging multi-view image inputs alongside multimodal intelligent guidance, Homie maintains consistent visual features and realistic physical interactions between subjects and objects across generated video frames.

MODEL1h ago

Microsoft releases Mage Flow 4B image model

Microsoft has released Mage Flow, an open-source 4-billion parameter model family designed for high-efficiency text-to-image synthesis and fine-grained editing. Combining a one-step latent tokenizer (Mage-VAE) with a Native-Resolution Multimodal Diffusion Transformer (NR-MMDiT), the MIT-licensed suite supports resolutions from 512 to 2048 pixels alongside sub-second Turbo variants.