We tested 4 speech-to-text engines on a 55-minute recording

We ran one 55-minute English recording through two cloud speech-to-text services and two models that run entirely on a phone, then scored every transcript against the same reference transcript. Cloud was faster and more accurate, but a 110-million-parameter model on the phone transcribed the whole recording offline in under three minutes at about 12% word error rate, which is good enough for notes.

Results

WER is word error rate: substituted, missing and extra words divided by the words in the reference. Lower is better. Tested October 8–9, 2026.

How we tested

• Recording: a 55-minute public company earnings call in English, 12.9 MB as M4A, with several speakers, numbers and financial terms.
• Reference: a reference transcript of 8,819 words.
• Phone: a Samsung Galaxy S24 FE, using the Audio to Text tool in the VoxaType app with the language left on automatic detection. The two on-device models ran with no network involved.
• Scoring: standard word error rate, once on the full text and once with all numbers removed from both sides, because some engines write "twenty-four" where the reference has "24", which counts as errors without being a mishearing.

What the numbers say

Cloud is still the most accurate and the fastest. Groq's whisper-large-v3-turbo finished the 55 minutes in 13 seconds at 9.6% WER. ElevenLabs Scribe v2 took 39 seconds; its raw 12.8% came mostly from spelling numbers out and keeping filler words, and with numbers excluded it was the most accurate of the four at 7.0%.

A small model on the phone is usable. Parakeet 110M took 171 seconds, roughly 20 times faster than real time, and scored 17.2%. About five points of that were numbers written as words: once we converted spoken numbers to digits on the phone, the same model scored 12.0% in 162 seconds. For notes and drafts that is workable, and the audio never leaves the device.

The smallest multilingual model breaks on long audio. Whisper Tiny took 424 seconds and scored 56.8%, mostly because it got stuck in repetition loops: 19 of them, covering 2,061 of its words, with single words repeated more than a hundred times. Detecting a loop and re-decoding that stretch in smaller pieces brought it to 40.2%, which is still far behind the others.

Short clips tell a different story

On short dictation clips recorded on the same phone, the ranking was less clear. Whisper large-v3 wrote "JGPT" for "ChatGPT" on every attempt, even with the word supplied as a hint, while Scribe v2 got it right unaided. Whisper also tends to produce a confident "Thank you" when given silence, so a dictation app has to check for real speech before inserting text. One long recording and a handful of clips are not a full benchmark; your audio, accent and vocabulary will move these numbers.

Choosing for your own use

• Quotes, interviews, anything you will publish: use a cloud engine and read the transcript before you rely on it.
• Private notes, offline use, or no budget: an on-device model in the Parakeet class is accurate enough and keeps the audio on your phone.
• Recordings over a few minutes: avoid the tiniest models, or use a tool that detects and repairs repetition loops.

The VoxaType Android app uses exactly this split: free dictation and recordings run on the phone, and Pro adds cloud transcription for the hardest audio. To try a recording in your browser instead, use Audio to Text.