Helpers for signed 16-bit little-endian interleaved PCM, the format MobMedia.audio_pcm/2 returns, plus the reference conversion the Android decoder implements and MobMedia.Engine.Fake uses.

Conversion

convert/5 maps channels, then resamples each channel:

• Channels. Same count: unchanged. To mono: the average of every channel (truncated). Mono to stereo: the sample on both sides. More than two channels to stereo: the first two (front left, front right).
• Downsampling (in_rate > out_rate): output frame k is the average of input frames floor(k * in / out) up to, not including, floor((k + 1) * in / out): a box filter, which is enough of a low-pass for speech recognition. A box is emitted once all of its input frames exist.
• Upsampling (in_rate < out_rate): output frame k sits at input position k * in / out and is linearly interpolated between its two neighbours (the last input frame stands alone), rounded half up. Emitted while that position is inside the input.

So n input frames give about n * out / in output frames (within one). iOS uses Core Audio's own converter instead, so its samples differ slightly; the length is the same within a few milliseconds.

This is integer arithmetic only, so a streaming implementation (the Kotlin bridge converts one decoder buffer at a time) produces exactly the same samples as this one.