Sample Rate Conversion

In this document

See the Wikipedia article Resampling (audio) for a generic definition of sample rate conversion, also known as "resampling." The remainder of this article describes resampling within Android.

Sample rate conversion is the process of changing a stream of discrete samples at one sample rate to another stream at another sample rate. A sample rate converter, or resampler, is a module that implements sample rate conversion. With respect to the resampler, the original stream is called the source signal, and the resampled stream is the sink signal.

Resamplers are used in several places in Android. For example, an MP3 file may be encoded at 44.1 kHz sample rate and needs to be played back on an Android device supporting 48 kHz audio internally. In that case, a resampler would be used to upsample the MP3 output audio from 44.1 kHz source sample rate to a 48 kHz sink sample rate used within the Android device.

The characteristics of a resampler can be expressed using metrics, including:

  • degree of preservation of the overall amplitude of the signal
  • degree of preservation of the frequency bandwidth of the signal, subject to limitations of the sink sample rate
  • overall latency through the resampler
  • consistent phase and group delay with respect to frequency
  • computational complexity, expressed in CPU cycles or power draw
  • permitted ratios of source and sink sample rates
  • ability to dynamically change sample rate ratios
  • which digital audio sample formats are supported

The ideal resampler would exactly preserve the source signal's amplitude and frequency bandwidth (subject to limitations of the sink sample rate), have minimal and consistent delay, have minimal computational complexity, permit arbitrary and dynamic conversion ratios, and support all common digital audio sample formats. In practice, ideal resamplers do not exist, and actual resamplers are a compromise among these characteristics. For example, the goals of ideal quality conflict with short delay and low complexity.

Android includes a variety of audio resamplers, so that appropriate compromises can be made depending on the application use case and load. Section Resampler implementations below lists the available resamplers, summarizes their characteristics, and identifies where they should typically be used.

Terminology

downsample
to resample, where sink sample rate < source sample rate
Nyquist frequency
The Nyquist frequency, equal to 1/2 of a given sample rate, is the maximum frequency component that can be represented by a discretized signal at that sample rate. For example, the human hearing range is typically assumed to extend up to approximately 20 kHz, and so a digital audio signal must have a sample rate of at least 40 kHz to represent that range. In practice, sample rates of 44.1 kHz and 48 kHz are commonly used, with Nyquist frequencies of 22.05 kHz and 24 kHz respectively. See the Wikipedia articles Nyquist frequency and Hearing range for more information.
resampler
synonym for sample rate converter
resampling
the process of converting sample rate
sample rate converter
a module that resamples
sink
the output of a resampler
source
the input to a resampler
upsample
to resample, where sink sample rate > source sample rate

Resampler implementations

Available resampler implementations change frequently, and may be customized by OEMs. As of Android 4.4, the default resamplers in descending order of signal distortion, and ascending order of computational complexity include:

  • linear
  • cubic
  • sinc with original coefficients
  • sinc with revised coefficients

In general, the sinc resamplers are more appropriate for higher-quality music playback, and the other resamplers should be reserved for cases where quality is less important (an example might be "key clicks" or similar).

The specific resampler implementation selected depends on the use case, load, and the value of system property af.resampler.quality. For details, consult the audio resampler source code in AudioFlinger.