> Complete local text-to-waveform speech synthesis under 10M parameters.
In case, like me, you hoped "complete" voice might mean both stt and tts. Not to speak poorly of it, just clarifying.
> English only, with one fixed male voice. This is not zero-shot voice cloning.
(And then a bunch of statements on limitations that I read as 'quality can be spotty but if you play with it it should be fine') But like. In <10M params I'm not judging:)
this is extremely encouraging for individuals/small companies being able to train pareto-frontier TTS models (specifically compute required to run vs quality of model output)
I think we need more neurons in the human brain than parameters in this model for speech. I wonder what it says about the human brain vs LLM efficiency.
This is amazing, the quality blow my mind for such small model! I just replaced my old onnx model with yours!
here my implementation with speech dispatcher and server: https://github.com/skorotkiewicz/inflect-speechd
thanks for shearing!
Couple highlights:
> Complete local text-to-waveform speech synthesis under 10M parameters.
In case, like me, you hoped "complete" voice might mean both stt and tts. Not to speak poorly of it, just clarifying.
> English only, with one fixed male voice. This is not zero-shot voice cloning.
(And then a bunch of statements on limitations that I read as 'quality can be spotty but if you play with it it should be fine') But like. In <10M params I'm not judging:)
When would “text-to-waveform speech synthesis” ever imply speech to text?
The HN title is "Inflect-Micro-v2: complete voice in 9.36M parameters".
The inflections are weird but this doesn't sound like a robot. Not bad!
This is impressive. I wish there were a voice clone option.
With so few parameters, I imagine a voice fine-tune might be readily tractable.
I keep seeing tts stories here. Is it just an interesting subset of the llm world, or is there a huge use case I’m somehow missing?
I just built my own voice assistant with my Pebble Time 2 watch and it uses a VPS hosted TTS (Piper) and Hermes agent.
I learned about all of these projects on HN at one point or another.
I use STT/TTS to interface with a local LLM for Home Assistant in my house.
I'm do so as well, i tried qwen3 omni 3 but it was ridiculously stupid, and i ended up with stt thinker and tts. kokoro for now
How heavy in inference on this? The model would easily fit on many microcontroller modules, I wonder if they could run it?
this is extremely encouraging for individuals/small companies being able to train pareto-frontier TTS models (specifically compute required to run vs quality of model output)
I think we need more neurons in the human brain than parameters in this model for speech. I wonder what it says about the human brain vs LLM efficiency.
I'd love to hear it but it seems your quota is exhausted.
wrote a web-demo, runs in your browser
https://inflect-tts.geronimo-labs.com
code: https://github.com/geronimi73/inflect-tts
Nice! Strangely, "Nano" sounds a lot better than "Micro" to me.
On my iPhone 14 Pro the page crashes after 2-3 plays. I wonder if it uses too much memory?
4 replies →
amazing quality for such small size!
Amazing quality for small size, but definitely not that enjoyable to listen to.
IMHO, its at about the same quality level of historic TTS tools.
I'm not sure which historic tools you mean, but to me this sounds much better than anything older than ten years ago.
I compared the macos Samantha just now and I guess the inflect-micro is marginally better...
Ivona „Joey“, „Amy“
amazing! was looking for something similar
Alternative title: Text to speech in 9.36M, English only.
[flagged]
[flagged]
[dead]
[dead]