AudioPreprocessing

Resample signed little-endian PCM (8, 16, 24 or 32 bit, despite the name) to targetRate using Java Sound.

The contract:

• '''Arguments are checked first.''' The target rate and the source's sample rate must be between MinSampleRate and MaxSampleRate Hz, the channel count between 1 and MaxChannels, and the bit depth a multiple of 8 from 8 to 32. Anything else is a Left(ValidationError) naming the field (targetRate, source.sampleRate, source.numChannels or source.bitDepth), and nothing reaches Java Sound.
• '''The output has exactly round(frames * targetRate / sourceRate) frames''' (half up), where frames is the number of whole frames in bytes. A trailing partial frame is ignored, as it is by toMono and trimSilence. Java Sound's converter delays the signal by its filter latency and appends the same number of frames as a tail; the length fit drops that tail, so the output is the source delayed by a fraction of a millisecond (at most 0.3 ms for 24 to 16, 16 to 24, 44.1 to 16, 16 to 48, 8 to 48, 48 to 8 and 22.05 to 44.1 kHz, measured on JDK 21) and the source's final fraction of a millisecond is not in it.
• '''Empty input gives empty output''', and so does input too short to make a single output frame.
• '''Equal rates return a copy of the whole frames''' without converting.
• '''An output larger than MaxOutputBytes (256 MiB) is a Left(ValidationError) on targetRate''', checked from the expected frame count before anything is allocated. The output is then written into one array of exactly the expected size.
• '''A converter that delivers fewer than the expected frames''' (by more than MaxShortfallFrames, or no frames at all) is a Left(ProcessingError); a shortfall within that is padded with silence.
• '''It cannot loop without making progress:''' reading stops at the end of the stream, at a read that returns no bytes, and once the expected number of bytes has been read, whichever comes first.

Compose multiple steps functionally

standardizeForSTT for generated audio; MP3 is a ValidationError, as it is not PCM.

Convert to mono by averaging channels (PCM16 little-endian).

Trim leading and trailing silence using a simple amplitude threshold on PCM16.