Bluetooth Headset Dictation Accuracy, Tested

Dictation loses a modest amount of accuracy through a Bluetooth headset's call link, and almost all of it comes from the narrowband path. On the wideband mSBC path, no engine got measurably worse. On the narrowband CVSD path, every engine's error rate rose, but the rise was measurable beyond chance only for Whisper large-v3-turbo. This page rests on 12 data points ¹: three engines, each run on the same 100 English recordings as recorded and after each of three voice paths.

Whisper loses the most on the narrowband link

Whisper large-v3-turbo scored 3.92 percent word error on the original recordings ². After the CVSD path every hands-free headset must support, it scored 5.18 percent ³. That is the largest rise of the three engines, and one of only two changes in the test that stayed larger than chance after correcting for all nine comparisons. The other is Whisper on plain 8 kHz audio with no codec, at 5.00 percent ⁴.

Apple SpeechTranscriber moved from 6.40 percent ⁵ to 7.43 percent ⁶ on CVSD. That rise is smaller than Whisper's. Its interval over the 100 recordings stays above zero on its own, but it does not survive the correction for nine comparisons, so we do not call it a measured difference.

Parakeet TDT v3 moved least: 4.11 percent ⁷ to 4.53 percent ⁸. Its interval includes zero, so this change cannot be told apart from noise.

Most of the narrowband loss is the missing high frequencies, not the codec

The 8 kHz, no codec rows remove the codec and keep only the bandwidth limit. They sit just below the CVSD rows for every engine. Whisper scored 5.00 percent without the codec ⁴ against 5.18 percent with it ³. Apple scored 7.15 percent ⁹ against 7.43 percent ⁶. Parakeet scored 4.39 percent ¹⁰ against 4.53 percent ⁸.

In all three engines the codec adds a little on top of the bandwidth cut, and the gaps are small compared with the cut itself. The practical reading: the damage is done when the link throws away speech above the narrowband ceiling, and CVSD's own distortion adds only a slight extra cost. This test does not examine which words fail.

The wideband path costs nothing you can measure

The mSBC path keeps the full frequency range of the original recordings. Whisper scored 3.69 percent ¹¹, lower than its original 3.92 percent ². Apple scored 5.93 percent ¹², also lower than its original 6.40 percent ⁵. Parakeet scored 4.20 percent ¹³, slightly higher than its original 4.11 percent ⁷.

Two engines improved and one worsened, all by small amounts. None of the three changes is a measured difference after correction. Apple's improvement has an interval that stays below zero by itself, but it is the same kind of result as its narrowband rise: not strong enough once nine comparisons are counted. Do not read it as wideband making Apple's engine more accurate. The honest summary is that wideband and the original recording are indistinguishable here.

The best engine depends on which link you have

Whisper has the lowest error on the original recordings (3.92 percent ², against 4.11 percent for Parakeet ⁷) and on the wideband path (3.69 percent ¹¹ against 4.20 percent ¹³). On the narrowband paths the order flips. Parakeet scored 4.53 percent on CVSD ⁸, below Whisper's 5.18 percent ³, and the same holds on plain 8 kHz audio.

Apple SpeechTranscriber has the highest error in every condition. Its best result, 5.93 percent on wideband ¹², is still above Whisper's and Parakeet's worst results. Its narrowband figure of 7.43 percent ⁶ sits well above the other two.

Where the numbers disagree is the ranking of the top two. The gap between Whisper and Parakeet on each path is small, and we have one test of 100 recordings, so we would not call the flip settled. The finding that holds is the size of the rise: Whisper's was the largest, Parakeet's the smallest.

What this test cannot tell you

The codecs ran in software on studio-clean read speech. A real headset adds its own microphone, noise suppression and radio dropouts, none of which are in these rows, and any of them could cost more accuracy than the codec does. Compare conditions and engines with each other rather than reading these percentages as what a headset in a room will give you. For how engines behave when the room itself is noisy, see our data on noise robustness.

The test also leaves out the newer LE Audio codec (LC3) and makers' own codecs, so it says nothing about headsets that use them. It covers English only. It uses one machine and one set of recordings, so the engine ranking may shift with other speakers, accents or vocabulary. For how these engines compare when run locally without any headset, see the on-device benchmark.

What to do if you dictate through a Bluetooth headset

Prefer a wideband link. Most of the loss we saw comes from the narrowband bandwidth cut, and the wideband path avoided it. If your headset and computer negotiate mSBC for calls, dictation through it should be close to dictating into the original audio. If your device falls back to CVSD, expect somewhat more errors.

Pick the engine for the link you have. On a narrowband link Parakeet had the lower error of the two open engines, and Apple's built-in engine had the highest. On wideband, Whisper led. If you cannot control the link, Parakeet is the steadier choice here, with the smallest rise.

Consider the microphone first. The headset's microphone and room noise are not part of this test and may matter more. A wired or built-in microphone avoids the call link entirely. Our guide to the dictation microphone covers the choices, and the overview of speech recognition explains how engines turn audio into text.

The size of every loss here is small next to the difference between engines in the original recordings. If you already dictate through a Bluetooth headset and are satisfied with the results, this data gives you no strong reason to stop.