Loquela
Research

How we measured Loquela Remote's speed

An hour of audio in under 9 seconds, and a full hour at 18,000 audio minutes an hour with every request transcribed. How both tests were run.

4 min readResearch checked 2026-10-27

Loquela Remote transcribes at 400× transcribe speed: an hour of audio in under 9 seconds. That number comes from one test. A second test asked a different question: what happens when hundreds of hours of audio arrive every hour from many people at once. Both ran on October 27, 2026, on the release of Loquela Remote that was live that day. This post explains how.

TestResult
A 60-minute file, one at a time7.08 to 8.66 seconds of server time in every run
SpeedOver 400× transcribe speed in every run
One hour at 18,000 audio minutes an hour18,350 requests, every one transcribed
Short clips during that hourHalf back within 0.23 seconds, 95 in 100 within 2.2 seconds

One hour of audio, one file at a time

This test measures how fast one recording becomes text when it has the server to itself.

  • The service: a copy of the released service, with nothing else running on the server.
  • The files: four 60-minute speech recordings in common formats, listed in the table below.
  • The runs: one warm-up request, then three runs of each file.
  • The timing: server time is the service's own record of each request. The files were sent from the server itself, so the upload took almost no time.
File (60 minutes)Server time, three runsSlowest, timed by the sender
MP3, 16 kHz mono7.08 to 7.15 s7.25 s
AAC, 48 kHz mono, 64 kbps7.54 to 7.58 s7.66 s
AAC, 44.1 kHz stereo, 128 kbps8.21 to 8.25 s8.42 s
MP3, 44.1 kHz stereo, 128 kbps8.56 to 8.66 s8.89 s

The slowest run, timed from the sending side, took 8.89 seconds. An hour is 3,600 seconds of audio, so that is about 405 times faster than playing the recording. Every run beat 400×, which is the figure the site uses.

Upload time is not in these numbers, because it depends on your connection. A 60-minute MP3 at 128 kbps is about 58 MB. Mono at 64 kbps is half that, so it reaches us sooner.

Many people at once: one hour at 18,000 audio minutes an hour

This test measures what the service does under heavy, mixed traffic. 18,000 audio minutes an hour is 300 hours of audio arriving every hour. We kept it up for a full hour.

  • The service: the same release, set up the way it runs live.
  • The senders: 16 test accounts sending recordings at random moments, the way unrelated customers do. They ran on the server itself, so the times below leave out the internet.
  • The recordings: short voice clips of 5 to 30 seconds (half WAV, half M4A), meetings of 10 to 60 minutes and long recordings of 60 to 120 minutes (both MP3). By minutes of audio, 15% were short clips, 45% meetings and 40% long recordings.
  • Speaker labels: turned on for about 3 in 10 meetings and long recordings.
  • The speech: English read aloud, from LibriSpeech, a public collection of audiobook recordings, cut and joined to each length.
RecordingRequestsTranscribedMedian time
Short clips, 5 to 30 s18,02018,0200.23 s
Meetings, 10 to 60 min24124112.2 s
Long recordings, 60 to 120 min898942.6 s

In total, 18,925 minutes of audio arrived during the hour, and all 18,350 requests came back with a transcript. None was turned away and none failed.

Short clips stayed quick the whole time. In every ten minutes of the hour, their median stayed between 0.20 and 0.29 seconds, and over the hour 95 in 100 came back within 2.2 seconds.

Meetings ran at a median of 140× transcribe speed and long recordings at 122×. That is slower than in the first test because many recordings share the server at once. Each one takes longer than it would alone, and far more audio gets done in the hour.

What this means for your recordings

  • A 30-minute meeting takes about 4 seconds of server time when it has the server to itself, plus the time to upload it.
  • A short voice clip usually comes back in well under a second.
  • Your total wait is the upload, the server time and the transcript coming back. The transcript is small, so it arrives almost at once.

When we measure again

These figures are from the release that was live on October 27, 2026. We will measure again after changes to how transcription runs, and update the site and this post when the numbers change.

Loquela Remote is a private speech-to-text API and MCP server for customers in the United States. It keeps no audio and no transcripts. See what happens to your audio, how other transcription APIs handle it and pricing.