Is 8 kHz audio good enough for speech recognition?

Mostly yes, with a small cost. In our test, 8 kHz audio raised the word error rate of every engine a little, and the rise was larger than chance only for Whisper large-v3-turbo. Parakeet TDT v3 and Apple SpeechTranscriber stayed within the range that cannot be told apart from no change. Speech recognition still works well at 8 kHz, but it is not free.

Cutting to 8 kHz removes every sound above 4 kHz. That is the bandwidth of a phone call and of the narrowband Bluetooth voice path.

Whisper lost the most when audio was cut to 8 kHz

On the original recordings Whisper made 3.92 percent word errors ¹. Cut to 8 kHz with no codec, it made 5.00 percent ². Over the 100 recordings that rise is larger than chance across all nine comparisons in the test.

Parakeet and Apple lost less than the test can resolve

Parakeet TDT v3 went from 4.11 percent ³ to 4.39 percent at 8 kHz ⁴. Apple SpeechTranscriber went from 6.40 percent ⁵ to 7.15 percent ⁶.

Both figures are higher, but the interval around each change, allowing for the test's nine comparisons, includes zero. The honest reading is that any loss is small for these two engines and that 100 recordings are too few to pin it down. Apple's change is the larger of the two, so a small loss is possible there.

The codec adds little on top of the bandwidth cut

The narrowband Bluetooth path adds the CVSD codec to the 8 kHz audio. For every engine it scored a little worse than plain 8 kHz audio:

• Whisper: 5.18 percent ⁷
• Parakeet: 4.53 percent ⁸
• Apple: 7.43 percent ⁹

For each engine, most of the total change from the original comes from the cut to 8 kHz, and the codec step is the smaller part. Bandwidth matters more than the codec.

Wideband audio avoids the loss

For comparison with 16 kHz audio over a Bluetooth path, the wideband (mSBC) path keeps that bandwidth. Whisper made 3.69 percent errors over it ¹⁰, Parakeet 4.20 percent ¹¹ and Apple 5.93 percent ¹². None differed measurably from the originals, so 16 kHz speech survived the wideband path intact. The full results across all four audio paths are in the chart and tables below.

The engines swap places at 8 kHz

At 16 kHz Whisper had the lowest error rate of the three engines. At 8 kHz Parakeet did, with Whisper second and Apple highest. That swap is consistent with Whisper losing more from the cut. It is a within-engine finding, not a proven gap between engines, because the test compared each engine with itself.

What these numbers do and do not measure

The test used 100 English recordings of clean read speech. The 8 kHz audio was made by cutting each recording in software, and each path started from the same recording at the same speech level. It did not use real headsets, so microphone quality, noise suppression and radio dropouts are missing. It also leaves out the newer LE Audio codec (LC3) and makers' own codecs. Real speech in a noisy room is harder, and low bandwidth may hurt more there. See our data on noise robustness.

What this means for choosing how to dictate

1. Prefer 16 kHz audio when you can. A built-in or wired microphone avoids the call link. Our guide to the dictation microphone covers the options.
2. Do not discard narrowband audio as useless. Recordings from phone calls and narrowband headsets are still usable, especially with an engine that held up better.
3. Choose the engine to match the audio. If your audio is often narrowband, Parakeet held up best here. Compare tools on the market page and in the on-device benchmark.

For the broader topic, read our overview of dictation, or open the full dataset.