Fermion ResearchFermion Research

Glossary/facts checked 13 November 2026

Real-time factor (RTF) and RTFx

Two ways of writing the same speed number, and the measurement details that decide whether two of them can be compared.

Phonon-2.5 benchmarks and downloadsOpen weights on Hugging Face.

The short answer

Real-time factor (RTF) is the time a system takes to process audio divided by the length of that audio: lower is faster, and anything below 1 keeps up with live speech. RTFx, the inverse real-time factor, is audio length divided by processing time: higher is faster.

"174x realtime" is an RTFx of 174, which is an RTF of about 0.0057. It means one hour of audio is transcribed in about 21 seconds (3,600 / 174 = 20.7).

RTF, RTFx and latency

The first two are throughput measures and convert into each other. Latency is a different thing.

MeasureFormulaBetter whenFor a 174x realtime systemAnswers
Real-time factor (RTF)processing time / audio durationLower0.0057Can it keep up with live audio?
RTFx (inverse RTF, "x realtime")audio duration / processing timeHigher174How much audio per second of compute?
Time per audio hour3,600 s / RTFxLowerabout 21 sHow long does a batch job take?
Latencytime from end of speech to final textLowernot implied by RTFxHow long does a user wait?

Reading published figures: Phonon-2 as an example

The figures below are Phonon-2's published numbers, converted. There are no published speed figures for Phonon-2.5. Note the last row: a batched figure is total throughput across many files at once, not the speed of one stream.

Phonon-2, publishedx realtime (RTFx)RTFOne hour of audio
Apple Neural Engine via Core ML, M5 MacBook Air6060.0017about 6 s
GPU of a base M5 MacBook Air via MLX1740.0057about 21 s
Eight AMD Zen 5 cores, CPU engine1430.0070about 25 s
Google Axion (Arm CPU)520.0192about 69 s
CPU of an M5 MacBook Air400.025090 s
NVIDIA H100, batched6,6800.00015about 0.5 s of compute per hour, across a batch

When two speed numbers are comparable

  • Same hardware: a GPU number and a laptop CPU number differ by orders of magnitude for the same model.
  • Same batching: one stream at a time and a large batch measure different things.
  • Same scope: some figures time only the model, others include loading the model, decoding the audio file and writing output.
  • Same audio: long files and short clips amortise overheads differently.

The Open ASR Leaderboard reports RTFx for every model it runs, measured by the maintainers on one hardware type (the repository names a single H200 for official runs). Leaderboard RTFx values compare models with each other on that machine; they do not predict speed on a laptop.

Questions

What is real-time factor in speech recognition?
Processing time divided by audio duration. An RTF of 0.5 means a minute of audio takes 30 seconds; below 1 the system keeps up with live audio.
What is RTFx?
The inverse real-time factor: audio duration divided by processing time. An RTFx of 1,000 means an hour of audio in 3.6 seconds.
What does 174x realtime mean?
174 seconds of audio are processed per second of compute, so an hour of audio takes about 21 seconds. That is Phonon-2's published figure on the GPU of a base M5 MacBook Air via MLX.
Is a high RTFx the same as low latency?
No. RTFx measures throughput; latency is how long a user waits after speaking. A batched server can have a very high RTFx and still return each result later than a single-stream system.
How fast is Phonon-2.5?
No speed figures for Phonon-2.5 are published. The published figures on this page are Phonon-2's.
Can I compare leaderboard RTFx with my laptop?
No. Leaderboard RTFx is measured on the maintainers' GPU; measure on your own hardware with your own audio.

Related

Sources and method

  • Speed figures: Phonon-2's published figures (/models/phonon-2/ and /research/phonon-2/), M5 MacBook Air, AMD Zen 5, Google Axion and H100 hosts as stated.
  • Open ASR Leaderboard: github.com/huggingface/open_asr_leaderboard README, read 2026-11-13.
  • Srivastav et al., Open ASR Leaderboard: Towards Reproducible and Transparent Multilingual and Long-Form Speech Recognition Evaluation, arXiv 2510.06961 (first version 2025-10-08), read 2026-11-13.

Canonical: https://www.fermionresearch.com/glossary/real-time-factor/