Skip to main content

Fermion ResearchAudio modelOpen weights

Phonon-2

PopularityTop group about equally popular, 8 models

51out of 100new calculation

based on two data groups, mostly search, range 30 to 71

Summary

Phonon-2 is a heavily compressed English ASR derivative of NVIDIA Parakeet TDT 0.6B. Quantized weights bring it below 900 MB and target local execution rather than claiming a new foundation model.

The model card reports 5.21% word error rate across seven English datasets and roughly 174 times real-time on an M5 MacBook Air. CC-BY-4.0 applies to this derivative.

GraduallyBench Audio

Phonon-2

Partially tested
  • The test coverage is incomplete.
  • The model’s language coverage was not fully checked.
  • A regular process exit is not attested.
  • Only the indicated language subset was tested.
Show test details

Speech recognition

Local · English

20 valid outputs from 20 attempted and 36 planned test slots.

Word error rate (WER)
2.45 %(12 / 490 reference words)English
Character error rate (CER)
0.85 %(22 / 2,580 reference characters)English
Silence cases with words in the final transcript
0 %(0 / 4 assessable silence cases)EnglishAfter Unicode normalization, word segments containing letters or digits count. Punctuation-only output does not count.

Speech recognition

Local · German

16 valid outputs from 16 attempted and 16 planned test slots.

Word error rate (WER)
31.09 %(171 / 550 reference words)German
Character error rate (CER)
10.36 %(338 / 3,262 reference characters)German

WER compares words and CER compares characters with the reference. The speech output was re-transcribed with GPT Transcribe. Its exact backend revision is unconfirmed. Read speech and digital silence do not cover spontaneous speech or real background noise. Error rates can exceed 100 %. Where used, the test references were not independently reviewed by humans. Test coverage and methods differ by model. Each metric stands on its own; there is no combined overall score across tests. Data compiled on 11/10/2026. How we test · Download all Decision and Audio results as JSON

All benchmark results

Every published result with its source, test setup and comparison models.

Show all 37 published benchmark results

Page 1 of 4

Measurements without matching peer values

These measurements have no matching peer values under the same test conditions. Their original values and sources remain available here.

Open ASR Leaderboard pipeline, seven-set mean WER (Show measurement, test conditions, and source)
Source value
5.15
Metric
Mean word error rate across seven English test sets
Unit
%
Benchmark version
Open ASR Leaderboard English short latest.csv, result commit d2c5b384 dated 2026-09-25; scoring code 6fec425b
Category
speech recognition
Direction
Lower is better
Source type
vendor-reported
Evaluator
FermionResearch/Phonon-2 model card, Hugging Face revision 88baeef8afde2d9f410a2d3337f689d632d2d53f
Status
active
Retrieved at
2026-10-20
Methodology
Vendor-reported result. The pinned model card identifies the separate 5.15 value as the leaderboard's own evaluation-pipeline result on the same seven English test sets. It cites english_short_latest.csv, commit d2c5b384 (2026-09-25), scoring code 6fec425b, kaldialign 0.12.0, compound merge, and greedy decoding. The evaluator CSV could not be read directly in this pass, so the value remains publisher-relayed.
Context
This row is publisher-relayed from the leaderboard's own evaluation pipeline. The exact underlying model-weight revision is not independently attested by the accessible leaderboard material. No aggregate sample denominator or confidence interval is stated for 5.15; neither is inferred. The direct result file was inaccessible and the current dataset page showed only a different model row.
Source FermionResearch/Phonon-2 model card, Hugging Face revision 88baeef8afde2d9f410a2d3337f689d632d2d53f
FluidAudio Phonon-2 LibriSpeech test-clean WER (Show measurement, test conditions, and source)
Source value
2.47
Metric
Corpus-level word error rate on LibriSpeech test-clean
Unit
%
Direction
Lower is better
Source type
community-reported
Evaluator
FluidAudio / FluidInference documentation
Status
active
Retrieved at
2026-10-22
Methodology
Community-reported result. FluidAudio's Core ML build of Fermion Research Phonon-2, English-only, measured with FluidAudio asr-benchmark on an M5 Pro running macOS 27 using the default Neural Engine. The source reports full LibriSpeech test sets, corpus-level WER, and RTFx as total audio duration divided by processing time.
Context
These are FluidAudio's own measurements of its Core ML implementation, not a general result for every Phonon-2 deployment. The source gives no immutable FluidAudio build or exact measurement date.
Source FluidAudio / FluidInference documentation
FluidAudio Phonon-2 LibriSpeech test-other WER (Show measurement, test conditions, and source)
Source value
4.62
Metric
Corpus-level word error rate on LibriSpeech test-other
Unit
%
Direction
Lower is better
Source type
community-reported
Evaluator
FluidAudio / FluidInference documentation
Status
active
Retrieved at
2026-10-22
Methodology
Community-reported result. FluidAudio's Core ML build of Fermion Research Phonon-2, English-only, measured with FluidAudio asr-benchmark on an M5 Pro running macOS 27 using the default Neural Engine. The source reports full LibriSpeech test sets, corpus-level WER, and RTFx as total audio duration divided by processing time.
Context
These are FluidAudio's own measurements of its Core ML implementation, not a general result for every Phonon-2 deployment. The source gives no immutable FluidAudio build or exact measurement date.
Source FluidAudio / FluidInference documentation
FluidAudio Phonon-2 Earnings-22 custom four-call 3,600-second file pause-cut WER (Show measurement, test conditions, and source)
Source value
16.7
Metric
Corpus-level word error rate on the concatenated Earnings-22 reference transcripts
Unit
%
Direction
Lower is better
Source type
community-reported
Evaluator
FluidAudio / FluidInference documentation
Status
active
Retrieved at
2026-10-22
Methodology
Community-reported result. FluidAudio transcribed the 3,600-second earnings22_top4_1h.wav file, formed from four concatenated Earnings-22 calls, on an M5 Pro with the default Neural Engine ladder encoder and pause-cut windows. The reference joins the Earnings-22 chunk transcripts; processing time excludes model loading. The page reports the best of two runs in one session.
Context
The page reports the best of two runs, not a repeated-run distribution. It does not publish an exact measurement date or immutable FluidAudio build; no sample count or confidence interval is inferred from the four concatenated calls.
Source FluidAudio / FluidInference documentation
average WER (Show measurement, test conditions, and source)
Source value
5.21
Metric
word error rate
Unit
%
Benchmark version
source version not stated
Category
audio
Direction
Lower is better
Source type
vendor-reported
Evaluator
Phonon-2 model card
Status
active
Retrieved at
2026-10-04
Methodology
Vendor-reported result. LS-clean 1.72, LS-other 3.92, AMI 9.37, Earnings22 6.96, GigaSpeech 8.35, SPGISpeech 3.70, VoxPopuli 2.46
Context
Source-specific observation; it does not establish a compatible cross-model cohort.
Source Phonon-2 model card
Phonon-2 English short-set WER: AMI (Show measurement, test conditions, and source)
Source value
9.37
Metric
WER
Unit
%
Benchmark version
Hugging Face Open ASR leaderboard English short-set snapshot 2026-09-25 (commit d2c5b384)
Category
audio
Direction
Lower is better
Sample
7805
Source type
vendor-reported
Evaluator
Fermion Research Phonon-2 model card
Status
active
Retrieved at
2026-10-16
Methodology
Vendor-reported result. Phonon-2 0.60B, English 16 kHz, greedy decoding, compound merging enabled; WER scored with kaldialign 0.12.0 over the full test set. Board snapshot: 2026-09-25, commit d2c5b384. The reported sample count is the exact evaluated test utterance count where the card publishes one.
Context
The model-card N is the exact full-test utterance count, not training data. No confidence interval is published. The card's 5.21 average already exists in prior audio research and is not repeated here.
Source Fermion Research Phonon-2 model card
Phonon-2 English short-set WER: Earnings-22 (Show measurement, test conditions, and source)
Source value
6.96
Metric
WER
Unit
%
Benchmark version
Hugging Face Open ASR leaderboard English short-set snapshot 2026-09-25 (commit d2c5b384)
Category
audio
Direction
Lower is better
Source type
vendor-reported
Evaluator
Fermion Research Phonon-2 model card
Status
active
Retrieved at
2026-10-16
Methodology
Vendor-reported result. Phonon-2 0.60B, English 16 kHz, greedy decoding, compound merging enabled; WER scored with kaldialign 0.12.0 over the full test set. Board snapshot: 2026-09-25, commit d2c5b384. The reported sample count is the exact evaluated test utterance count where the card publishes one.
Context
The model card says Earnings-22 comprises six sessions evaluated in chunks and publishes no chunk or utterance denominator. No sample count or confidence interval is inferred. The card's 5.21 average already exists in prior audio research and is not repeated here.
Source Fermion Research Phonon-2 model card
Phonon-2 English short-set WER: GigaSpeech (Show measurement, test conditions, and source)
Source value
8.35
Metric
WER
Unit
%
Benchmark version
Hugging Face Open ASR leaderboard English short-set snapshot 2026-09-25 (commit d2c5b384)
Category
audio
Direction
Lower is better
Sample
18768
Source type
vendor-reported
Evaluator
Fermion Research Phonon-2 model card
Status
active
Retrieved at
2026-10-16
Methodology
Vendor-reported result. Phonon-2 0.60B, English 16 kHz, greedy decoding, compound merging enabled; WER scored with kaldialign 0.12.0 over the full test set. Board snapshot: 2026-09-25, commit d2c5b384. The reported sample count is the exact evaluated test utterance count where the card publishes one.
Context
The model-card N is the exact full-test utterance count, not training data. No confidence interval is published. The card's 5.21 average already exists in prior audio research and is not repeated here.
Source Fermion Research Phonon-2 model card
Phonon-2 English short-set WER: LibriSpeech test-clean (Show measurement, test conditions, and source)
Source value
1.72
Metric
WER
Unit
%
Benchmark version
Hugging Face Open ASR leaderboard English short-set snapshot 2026-09-25 (commit d2c5b384)
Category
audio
Direction
Lower is better
Sample
2620
Source type
vendor-reported
Evaluator
Fermion Research Phonon-2 model card
Status
superseded
Retrieved at
2026-10-16
Methodology
Vendor-reported result. Phonon-2 0.60B, English 16 kHz, greedy decoding, compound merging enabled; WER scored with kaldialign 0.12.0 over the full test set. Board snapshot: 2026-09-25, commit d2c5b384. The reported sample count is the exact evaluated test utterance count where the card publishes one.
Context
The model-card N is the exact full-test utterance count, not training data. No confidence interval is published. The card's 5.21 average already exists in prior audio research and is not repeated here.
Source Fermion Research Phonon-2 model card
Phonon-2 English short-set WER: LibriSpeech test-other (Show measurement, test conditions, and source)
Source value
3.92
Metric
WER
Unit
%
Benchmark version
Hugging Face Open ASR leaderboard English short-set snapshot 2026-09-25 (commit d2c5b384)
Category
audio
Direction
Lower is better
Sample
2939
Source type
vendor-reported
Evaluator
Fermion Research Phonon-2 model card
Status
active
Retrieved at
2026-10-16
Methodology
Vendor-reported result. Phonon-2 0.60B, English 16 kHz, greedy decoding, compound merging enabled; WER scored with kaldialign 0.12.0 over the full test set. Board snapshot: 2026-09-25, commit d2c5b384. The reported sample count is the exact evaluated test utterance count where the card publishes one.
Context
The model-card N is the exact full-test utterance count, not training data. No confidence interval is published. The card's 5.21 average already exists in prior audio research and is not repeated here.
Source Fermion Research Phonon-2 model card
Phonon-2 English short-set WER: SPGISpeech (Show measurement, test conditions, and source)
Source value
3.7
Metric
WER
Unit
%
Benchmark version
Hugging Face Open ASR leaderboard English short-set snapshot 2026-09-25 (commit d2c5b384)
Category
audio
Direction
Lower is better
Sample
39341
Source type
vendor-reported
Evaluator
Fermion Research Phonon-2 model card
Status
active
Retrieved at
2026-10-16
Methodology
Vendor-reported result. Phonon-2 0.60B, English 16 kHz, greedy decoding, compound merging enabled; WER scored with kaldialign 0.12.0 over the full test set. Board snapshot: 2026-09-25, commit d2c5b384. The reported sample count is the exact evaluated test utterance count where the card publishes one.
Context
The model-card N is the exact full-test utterance count, not training data. No confidence interval is published. The card's 5.21 average already exists in prior audio research and is not repeated here.
Source Fermion Research Phonon-2 model card
Phonon-2 English short-set WER: VoxPopuli (Show measurement, test conditions, and source)
Source value
2.46
Metric
WER
Unit
%
Benchmark version
Hugging Face Open ASR leaderboard English short-set snapshot 2026-09-25 (commit d2c5b384)
Category
audio
Direction
Lower is better
Sample
628
Source type
vendor-reported
Evaluator
Fermion Research Phonon-2 model card
Status
active
Retrieved at
2026-10-16
Methodology
Vendor-reported result. Phonon-2 0.60B, English 16 kHz, greedy decoding, compound merging enabled; WER scored with kaldialign 0.12.0 over the full test set. Board snapshot: 2026-09-25, commit d2c5b384. The reported sample count is the exact evaluated test utterance count where the card publishes one.
Context
The model-card N is the exact full-test utterance count, not training data. No confidence interval is published. The card's 5.21 average already exists in prior audio research and is not repeated here.
Source Fermion Research Phonon-2 model card

Page 1 of 4

Popularity compared

How often the model is searched for, used and mentioned, compared with every model of its category.

Popularity index

0 to 100, higher is more popular

20 of 54 models

As of November 2026 · Source: gradually.ai popularity index

Over time

Weekly values from October 18, 2026 to November 20, 2026

Phonon-2Gemini 3.5 Transcribe 52Fish Audio S2 Pro 52Nova-3 51MiniMax-Music3 50

Method changes in the timeline, a jump at these points is no market event:

  1. 1November 1, 2026: Value only with search or usage data, 30-day attention
  2. 2November 1, 2026: Method version 4 (common scale, shrinkage)
  3. 3November 9, 2026: New topics measurement V20 (daily quorum, short forms)

What is gaining popularity right now, across models, vendors, harnesses and AI subscriptions, is collected in All AI trends.

All specifications

Architecture

Model class
Compressed speech recognition
Source: Model class
Checkpoint
FermionResearch/Phonon-2
Source: Checkpoint

Input and output

Input
English audio
Source: Input
Output
Transcript
Source: Output

Watermarking and provenance

AI watermarking refers to invisible signals embedded in the output. Visible marks and provenance metadata are listed separately. Unknown means there is insufficient evidence for that particular field.

AI watermarking
Unknown
no source
Watermarking technique
Not known
no source

Licence and access

Access
Local weights
Source: Access
License
CC-BY-4.0
Source: License
Shutdown
None announced
no source

Further details

Derived from
Quantized derivative of nvidia/parakeet-tdt-0.6b-v3, not a new foundation model
Source: Derived from
Compression
About 2.1 bits per weight and under 900 MB
Source: Compression
Reported WER
5.21% across seven English datasets
Source: Reported WER

Sources and data date

Show all 17 sources

Benchmark sources at a glance

Which independent evaluators tested this model. This is coverage, not a quality ranking.

No active result from the reviewed external evaluators can currently be attributed to this model. Other source evidence remains available below.

Without vendor claims and historical or disputed results; those stay labelled in the single values.

TypeEvidence and data date
Research statusResearch date October 4, 2026. 13 source URLs checked. This documents the inspected sources, not an exhaustive inventory of every publication.
Not published

confidence-intervals

Note

The model card reports corpus WER point estimates without confidence intervals.

Additional sourcePhonon-2 model card (retrieved October 3, 2026; October 4, 2026) · Phonon-2 model card · Editorial description reviewed October 3, 2026
Additional sourceFermion Research Phonon-2 model card (retrieved October 16, 2026; October 20, 2026)
Additional sourceFermion Phonon-2 research page (retrieved October 19, 2026)
Additional sourceFermion Research Phonon-2 model page (retrieved October 21, 2026)
Additional sourceFluidAudio / FluidInference documentation (retrieved October 22, 2026)
Additional sourceHugging Face Open ASR Leaderboard (retrieved October 20, 2026; October 28, 2026)
Additional sourceModel metadata (retrieved October 16, 2026)
Additional sourceModel metadata (retrieved October 20, 2026)12
Additional sourceModel metadata (retrieved October 28, 2026)12
Additional sourceOpenRouter: Popularity index (derived), CC BY 4.0
Additional sourceVercel AI Gateway: Popularity index (derived), CC BY 4.0
Additional sourceOllama: cumulative downloads for the popularity index
Additional sourceReplicate: cumulative runs for the popularity index
Additional sourceGoogle Trending Now: limited feed selection for Germany and the US
Additional sourceCreative Commons Attribution 4.0