Sample rate
In one sentence
Sample rate is how many times per second audio is measured when it is digitized, and it sets a hard ceiling on which frequencies survive to reach the recognizer.
Definition
Sample rate is how many times a second sound is measured when it is turned into digital data. Measure more often and you capture more of the sound.
The rate sets a hard ceiling on which frequencies can be recorded, so it decides how much information ever reaches the recognizer.
The governing principle is that a sample rate can only represent frequencies up to half its value. This is why the numbers matter, and why upsampling cannot recover what was never captured in the first place.
The rates that appear in voice work
- 8 kHz, the telephony standard. Represents frequencies to 4 kHz, which is why phone calls sound the way they do.
- 16 kHz, the common standard for speech recognition. Represents to 8 kHz, which covers the speech-relevant range adequately.
- 24 kHz, frequently used for speech synthesis output, giving better quality on fricatives and sibilants.
- 44.1 and 48 kHz, the music and general-audio standards. More than speech requires, and usually downsampled before recognition.
Why the ceiling is absolute
- Frequencies above half the sample rate are not merely quietened; they are absent, and they can alias into the captured range as distortion if they are not filtered out first.
- Certain consonant distinctions, particularly among fricatives, carry energy in the higher frequencies that 8 kHz sampling discards.
- This is the technical explanation for why the same recognizer performs worse on telephony audio than on browser audio.
The upsampling trap
- Converting 8 kHz audio to 16 kHz produces a file in the right format and with no additional information.
- A model expecting 16 kHz input will process it without error and perform as though on narrowband audio, because it is.
- This is a common and invisible bug: everything works, and accuracy is quietly worse than it should be.
Practical guidance
- Match the sample rate the model expects. Feeding a mismatched rate produces either an error or, worse, plausible-looking garbage.
- Capture at the highest rate available and downsample if needed, rather than the reverse.
- Verify the actual rate in the media path rather than trusting configuration, since transcoding hops can change it silently.
Common misconception
That sample rate is a quality preference like image resolution. It is a hard information ceiling. Higher rates than the model needs waste bandwidth; lower rates destroy information permanently.
Why it matters commercially
Sample rate is one of the concrete technical reasons a browser-based voice agent has an accuracy advantage over a telephony one, and it is verifiable rather than a marketing claim.
In voice specifically
Typed text carries no audio, so there is no sample rate and no frequency ceiling underneath it. In speech the capture rate is fixed before the recognizer sees a thing, and nothing downstream can add back a frequency the microphone path already discarded.
Where AsqVox fits
Browser audio is typically captured at wideband rates, which gives the recognizer materially more information than the 8 kHz telephony path provides. This is a specification fact rather than a performance claim.
Visual
The ceiling you cannot raise later
Half the sample rate is the ceiling. Nothing recovers what was never captured.
Statistics
Every figure carries its source and year. Vendor numbers are labelled as vendor numbers, and where no reliable figure exists this page says so rather than borrowing one.
A sample rate can represent frequencies only up to half its value, which is why 8 kHz telephony audio carries content to about 4 kHz.
Half the sample rateindustry rangeNyquist-Shannon sampling theorem, 2026 - A mathematical ceiling, not a tuning choice. Frequencies above half the rate are absent, and can alias into the captured range as distortion if they are not filtered out first.
PSTN voice channels conventionally carry roughly 300 to 3,400 Hz, digitized at 8 kHz sampling.
300 to 3,400 Hzindustry rangePSTN telephony convention, 2026 - The narrowband voice band chosen decades ago to conserve bandwidth. It is why phone audio sounds the way it does.
G.711, the standard telephony codec, operates at 8 kHz sampling and 64 kbps.
8 kHz, 64 kbpsindustry rangeITU-T G.711, 2026 - Names the standard so the figure can be checked. The 8 kHz sampling is what caps telephony audio at roughly 4 kHz.
Opus supports sampling rates from narrowband up to 48 kHz fullband and adapts dynamically.
Up to 48 kHzindustry rangeOpus codec (RFC 6716), 2026 - A modern codec that is not tied to the 8 kHz telephony ceiling, which is why browser audio can be wideband.
16 kHz is the common input expectation for speech recognition models, representing frequencies to 8 kHz.
16 kHzindustry rangeSpeech recognition engineering convention, 2026 - Covers the speech-relevant range adequately. Feeding a model a rate it does not expect produces either an error or plausible-looking garbage.
Whisper Large-v3 records around 2.7 percent word error rate on clean benchmark audio and roughly 8 to 12 percent on real-world English, and narrowband telephony audio sits at the harder end of that range.
2.7% clean, 8 to 12% real worldindependentWhisper Large-v3 on LibriSpeech test-clean, against reported real-world English performance, 2026 - The benchmark half is a published result; the real-world half is a reported band. Narrowband capture is one of the conditions that pushes accuracy toward the worse end.
No reliable independent benchmark isolates recognition accuracy by sample rate for identical audio and models. The direction of the effect is well established; the magnitude is not publicly quantified.
-no reliable figureEveryone agrees narrowband is worse. Nobody has published a clean by-how-much for the same recording captured at different rates.
Examples
In practice
An integration passes telephony audio to a recognizer that expects 16 kHz. A resampling step upsamples from 8 kHz, so the format is correct and no error occurs. Accuracy is materially worse than on browser sessions using the same recognizer, and the cause is invisible in the logs because nothing failed. Auditing the media path for the original capture rate is the only way to find it.
The everyday version
Sample rate is how finely sound is measured when it is turned into data. Phone calls use a low setting chosen decades ago to save bandwidth, and it permanently throws away part of the sound. A browser captures much more. That is one concrete reason the same software understands people better through a website than through a phone line.
Usage
Who says it
- Audio engineers use the term routinely and precisely.
- It is rarely a business word, though it underpins claims about audio quality.
Where it turns up
- On a spec sheet it sits next to supported codecs, audio input requirements, transcoding behavior and bandwidth.
Common misuse
- Treating upsampling as a quality improvement.
- Assuming a correct file format implies correct information content.
- Trusting configured rates rather than verifying the actual rate in the media path.
Questions people ask
What sample rate does speech recognition need?
16 kHz is the common expectation. It represents frequencies to 8 kHz, which covers the speech range adequately. Match whatever rate the model expects: feeding it a mismatched rate produces either an error or plausible-looking garbage, so the rate is not something to guess at.
Does upsampling 8 kHz audio to 16 kHz improve accuracy?
No. It gives you the right file format and not one new piece of information. A 16 kHz model processes the upsampled file without error and performs as though on narrowband audio, because it is. This is a common and invisible bug: everything works and accuracy is quietly worse than it should be.
Why does the same voice agent understand people better on a website than on a phone call?
Telephony sampling is 8 kHz, which caps content at about 4 kHz and discards the consonant energy above it. Browser audio is captured at wideband rates, so the recognizer receives materially more information. It is a verifiable specification fact rather than a marketing claim.
What is the highest frequency a sample rate can capture?
Half the sample rate. Frequencies above that are absent, not merely attenuated, and they can alias into the captured range as distortion if they are not filtered out first. That is why the number is a hard ceiling rather than a quality preference.
Last reviewed 31 July 2026. Written and reviewed by Dhruv Dholakia, founder of AsqVox.