One open 9B audio model for speech recognition, understanding, TTS and editing
Emerged finds new products worth a look, daily — what’s new today
nano-edge-tts
Kokoro's voices with a 7 MB decoder in place of the original 213 MB one
Is this yours? Claim itIf Kokoro sounded right but was too heavy to ship inside your app, this cuts the whole model to 30M parameters with near-identical measured quality.
nano-edge-tts swaps the waveform decoder in Kokoro-82M for a 1.75M-parameter one. Pronunciation, timing, pitch and the 28 English voices are untouched, so it drops into an existing Kokoro pipeline. On one Apple M4 CPU thread the decoder ran 33 times faster than Kokoro's and the full path from phonemes to audio 11 times faster. Whisper word error and UTMOS scores land close to the original. Mike Chumba is candid about limits: English only, PyTorch only, timed on a laptop, and a faint machine tone on a few voices.
More like this
Fast, low-cost speech APIs built on Qwen3 voice models
An npm library for local speech-to-text on web pages, no server or API key
An open 4B speech model that listens and talks at once, in nine languages
Hold a key, speak, and the words land wherever the cursor is
A Mac menu-bar reader that speaks any text with the current word highlighted