audio sample rate is the number of samples used to represent one second of audio in each channel, measured in hertz (hz) or kilohertz (khz). a 24 khz recording contains 24,000 samples per second per channel. the rate affects the frequencies the audio can represent and how its samples are played.

how does audio sample rate work?

digital audio represents a waveform as a sequence of values. each sample describes the waveform's amplitude at one point in time. the sample rate tells a player how much time belongs between those values.

at 24 khz, 24,000 samples make up one second. at 8 khz, one second contains 8,000 samples per channel. these rates describe the audio representation, not how many words the speaker says each second.

sample rate also limits the frequencies that can be represented. the theoretical upper boundary is half the rate, with practical systems leaving room for filtering. mdn's digital audio guide explains this relationship and why lower-rate audio carries a narrower frequency range.

changing the rate requires resampling. this calculates samples for the new rate and uses filtering to manage the frequency range. increasing the rate of an existing recording doesn't restore high-frequency detail that the source didn't contain.

audio sample rate example

for example, an appointment-booking app could use rumik silk to stream a spoken confirmation. silk's default websocket output is raw 24 khz mono pcm, so a one-second portion contains 24,000 samples in one channel.

suppose the app mistakenly tells its player that those samples are 8 khz audio:

24,000 samples / 24,000 samples per second = 1 second
24,000 samples / 8,000 samples per second = 3 seconds

24,000 samples / 24,000 samples per second = 1 second
24,000 samples / 8,000 samples per second = 3 seconds

those samples now take three times as long to play, slowing the voice and lowering its pitch. the playback setting has changed how the audio is heard, even though the voice model received the same instructions.

if the receiving system needs 8 khz audio, the app needs a correctly converted signal. resampling one second from 24 khz to 8 khz produces 8,000 samples for that second. changing a label alone leaves the original 24,000 samples in place.

the distinction is important for raw audio because it doesn't carry a wav header telling the player how to interpret it. the app has to supply the correct format information. silk's audio-format guide documents the rate, channel count, and encoding for each output.

sample rate vs bit depth and bitrate

these values describe different parts of digital audio:

for uncompressed linear pcm, bitrate follows from the other values:

24,000 samples per second × 16 bits × 1 channel
= 384,000 bits per second

24,000 samples per second × 16 bits × 1 channel
= 384,000 bits per second

compressed formats use a codec to reduce the data required. two recordings can share a sample rate while having different bitrates. a bitrate value alone doesn't tell you the sample rate or how natural the generated voice sounds.

why does audio sample rate matter for voice agents?

the output from a tts service needs to match what the receiving player or phone system accepts. a mismatch can change playback speed or require conversion before the caller hears the reply.

silk's default http response is a 24 khz mono wav file. its mulaw and alaw output options use 8 khz mono audio for telephony. the format choice determines these documented output properties, so check the destination's requirements when selecting it.

a higher sample rate alone doesn't establish speech quality. pronunciation, noise, and the model's performance still affect what the listener hears. converting telephone audio to a higher rate won't recover detail already lost from the source.

in a streaming tts application, the sample rate also helps the player interpret each incoming chunk. the player needs consistent format information throughout playback. sample rate doesn't measure tts latency, although decoding, conversion, and buffering can add to the wait before speech begins.