Phonon-2 targets compact English speech recognition. Its compressed file size is worth examining, but hardware planning must account for the expanded execution representation and the complete transcription pipeline.
What the research and docs establish
Fermion reports a 164 MB download, 178 MB on disk and English recognition from 16 kHz mono audio. Weights are CC-BY-4.0, derived from NVIDIA Parakeet TDT 0.6B v3. Its research page explains expansion into a 16-bit encoder on Apple and NVIDIA GPUs, with a different CPU execution path. Reported speed excludes model loading unless specified. The provider’s benchmark averages and speed comparisons remain author-reported, not our measurements.
Compression is one boundary, execution another
Our recommended memory sheet separates download bytes, installed artifact bytes, loaded tensors, audio buffers and peak process memory. Compare those values at the same clip length and batch setting. A model that downloads quickly can still need more memory after unpacking. Do not buy a GPU from the compressed file headline; first identify the engine and execution representation used by your intended installation, then measure a representative job on available hardware.
Choose the right speech workload
For an English meeting trial, collect consented examples containing cross-talk, distant microphones and specialized names. Write reference transcripts and mark words whose errors change the decision: amounts, dates, negation and speaker intent. Our recommendation is to review those separately from average word error. A model may perform well on clean narration and still struggle on your meeting. Keep the recording environment and segmentation policy attached to each result so the comparison remains interpretable.
Do not mix batch throughput with live response
A batch pipeline processes many recordings; an interactive dictation system needs a prompt response to one speaker. Our suggested timing record includes model load, preprocessing, recognition and final text availability, with concurrent streams stated explicitly. Report the cold-start and warmed-up scenarios separately. If a published speed uses many simultaneous streams, it cannot be used unchanged to promise a single user’s latency. We have not reproduced any of the reported hardware figures.
Review the local application around the model
For a private transcription workflow, inspect where the application stores audio and transcripts and whether any optional service receives them. A local model does not by itself settle retention or access policy. Preserve attribution and check dependency terms when redistributing a package. Compare Whistle on the same recordings only where language and clip handling match; small download sizes are useful context, but accepted transcripts and total operational effort decide which workflow is appropriate. Keep corrected transcripts linked to the original recording so later summaries can be checked against the words actually spoken.
Frequently asked questions
Is Phonon-2 a multilingual speech model in these sources?
The reviewed release and documentation describe English recognition; do not infer support for other languages.
Why distinguish download size from runtime memory?
Execution can unpack or convert weights and needs buffers and other application components.
What makes a speech-speed comparison fair?
State hardware, clip lengths, engine, batch/concurrency and whether loading and preprocessing are included.
Sources & further reading
Primary sources checked Oct 11, 2026. Vendor statements are attributed; editorial advice is our own.
- 1Introducing Phonon-2 ↗Fermion Research
- 2Speech models documentation ↗Fermion Research
Help us keep this useful. Check contact availability →
