Does a Bluetooth headset lower dictation accuracy?

It depends on which Bluetooth voice path your headset uses. In our test, the wideband path (mSBC) cost no engine a measurable amount of accuracy. The narrowband path (CVSD), the one every hands-free headset must support, raised the word error rate measurably for one of the three engines, Whisper, and by a borderline amount for Apple SpeechTranscriber.

Whisper large-v3-turbo made 3.92 percent word errors on the original recordings ¹. After the narrowband path it made 5.18 percent ². Apple SpeechTranscriber went from 6.40 percent ³ to 7.43 percent ⁴.

The wideband path changed nothing measurable

Wideband audio keeps the 16 kHz sampling of the original recordings, so little of the speech is thrown away. Whisper made 3.69 percent errors through it ⁵. Parakeet TDT v3 made 4.20 percent ⁶, against 4.11 percent as recorded ⁷. Apple SpeechTranscriber made 5.93 percent ⁸.

Two engines read slightly lower than their originals and one slightly higher. That scatter is what noise looks like: once all nine comparisons in the test are taken into account, none of the three wideband changes is a measured difference.

The narrowband path hurt Whisper, possibly Apple, and not clearly Parakeet

The narrowband path squeezes speech into 8 kHz audio, which removes every sound above 4 kHz. Whisper's rise is larger than chance across all nine comparisons in the test. Apple SpeechTranscriber's rise is slightly smaller in points, and its interval across the nine comparisons includes zero, so we call it borderline rather than measured. Parakeet made 4.53 percent errors ⁹, higher than its original, but the interval around that change includes zero, so we cannot call it a loss.

In points, Whisper's rise was a little larger than Apple's. Relative to where each started, Whisper's rise was also larger, because Apple began from a higher error rate.

The ranking depends on the path

As recorded and over wideband, Whisper had the lowest error rate, Parakeet came next, and Apple SpeechTranscriber was highest. Over the narrowband path the order of the first two flipped: Parakeet made 4.53 percent errors and Whisper 5.18 percent. Each engine is compared with itself in this test, so read the flip as a hint that Parakeet holds up better on narrowband audio, not as a proven gap between engines.

The tables below list each engine across every path, from the full set of 12 data points ¹⁰.

What these numbers do and do not measure

Each engine heard the same 100 English recordings of read speech. Every Bluetooth path started from the same recording at the same speech level, and the codecs were run in software. That isolates the effect of the voice path.

It leaves out almost everything a real headset adds: its microphone, its noise suppression and radio dropouts. It also leaves out the newer LE Audio codec (LC3) and makers' own codecs. The recordings are clean studio speech, so absolute error rates in a room, on a train or with a weak signal will be higher. Use the numbers to compare paths and engines, not to predict your own error rate. For the effect of background noise, see our data on noise robustness.

What to look for in a headset

1. Check for wideband voice. Look for mSBC or "wideband speech" support in the headset's specifications. In this test it cost nothing measurable, while the fallback narrowband path did.
2. Expect narrowband in some situations. Narrowband is the baseline every hands-free headset must support, and a device can fall back to it. If errors jump when you connect a headset, that is a likely cause.
3. Consider the microphone you already have. A built-in or wired microphone sends 16 kHz audio straight to the engine with no call link in between. See our guide to the dictation microphone.
4. Pick the engine for the path. If you must dictate over narrowband, the data hints that Parakeet is the steadier choice. Compare tools on the market page.

For the wider picture, start with our overview of dictation, see how engines compare in the on-device benchmark, or open the full dataset.