Add Phonon-2 model

[keeandev]

https://www.fermionresearch.com/research/phonon-2/

✝️ Jesus loves you so much, that He chose to die for your sins on the cross.

📖 John 3:16 - “For God so loved the world, that He gave His only begotten Son, that whoever believes in Him shall not perish, but have eternal life.

[Mudxiej]

Adding technical context and real-world numbers on why this model is uniquely suited for Handy:

What makes Phonon-2 different:

1. Size & Footprint: Only 164 MB download (5-value ternary quantization at ~2.1 bits per weight). Drastically lighter on RAM and disk than Whisper large-v3 (3 GB) or Parakeet v3 (2.5 GB).
2. Architecture: It is a direct derivative of NVIDIA's parakeet-tdt-0.6b-v3 (weights under CC-BY-4.0).
3. Accuracy: Averages 5.21% WER across Open ASR Leaderboard English benchmarks (beating Whisper large-v3-turbo at 6.58% and matching its 2.5 GB full-precision teacher).
4. Zero Silence Hallucination: Because it uses a Token-and-Duration Transducer (TDT) decoder rather than an autoregressive Seq2Seq model like Whisper, it emits duration skips on dead air. It cannot get trapped in repetitive hallucination loops when pausing mid-sentence.
5. CPU Throughput: 10x-25x realtime on baseline CPUs, >140x on modern multi-core x86/ARM, and ~170x on Apple Silicon.

How this fits Handy:

• Handy already implemented parakeet.cpp / transcribe.cpp. The TDT decoder lattice is already familiar to the engine stack.
• Alternatively, if native C++ tensor quantization takes time, adding support for a local OpenAI-compatible STT endpoint (/v1/audio/transcriptions) in Handy's settings would allow running fermion serve phonon-2 locally right now with zero backend complexity.

Weights & reference repo:

• https://huggingface.co/FermionResearch/Phonon-2
• https://github.com/fermionresearch/phonon

[cjpais]

Please use your human brain to write. Not AI. It's effectively spam otherwise.

I would rather just a +1 than bullshit spam

[Mudxiej]

ok

[vanshdoingcode007]

Sir Please add a phonon 2 model plese and also parakeet redux also

• Phonon-2: https://huggingface.co/FermionResearch/Phonon-2
• Parakeet Redux: https://huggingface.co/suryatmodulus/parakeet-redux

Both look interesting for local speech-to-text. Could support for these models be considered in a future release? It would be great to have them available directly in Handy.

Thanks for working on this project!

[atesahmet0]

Check this out #213