v0.2.1

This is the first published release of the 0.2 series.

Upgrading from 0.1

0.1	0.2
AudioPipelineCoordinator::new_with_defaults()	Pipeline::new()?
AudioPipelineCoordinator::new(Arc::new(vad), chunker, dc, nr)	Pipeline::with_components(vad, chunker, dc, nr)?
coordinator.process_frame(&bytes)	pipeline.process_bytes(&bytes) (&mut self; .chunks holds the cleaned chunks)
VadDetector::new(config, metrics)	VadDetector::new(config) or VadDetector::with_metrics(config, metrics)
VADStats	vad::VadStats
VadConfig { activation_margin, release_margin, energy_weight, flux_weight, .. }	VadConfig { activation_ratio, release_ratio, .. }
chunk.overlap_next	the next chunk's overlap_prev, or chunk.samples_with_context()
AudioFormat::WebM	AudioFormat::Vorbis or AudioFormat::Opus
AudioBuffer, AudioChunk, the fixtures feature	removed

Error and AudioFormat are #[non_exhaustive], so a match on them
needs a wildcard arm. Opus needs features = ["opus"] (Rust 1.95).

Added

• Decoding of FLAC, ALAC, MP3, AAC (MP4/M4A), Ogg/WebM Vorbis, AIFF, CAF,
  and ADPCM through Symphonia 0.6, plus 8-bit, 32-bit, float, mu-law, and
  A-law WAV. Any channel count is downmixed to mono.
• AudioFormatConverter::convert_file for reading from a path.
• Pipeline::process_samples for already-decoded 16 kHz audio, and
  Pipeline::reset.
• VadConfig::min_speech_energy to tune the absolute level below which a
  frame is never speech.
• VadDetector::with_metrics for attaching a metrics collector.
• Pipeline::with_max_speech_hold (default 2 s): long uninterrupted speech
  is emitted in chunks while it is still going instead of after it ends.
• ProcessedChunk::samples_with_context: the previous chunk's tail followed
  by this chunk's samples. In the pipeline, the context also carries across
  process_samples calls.
• VadDetector::flush to analyze the last partial frame and close a segment
  that runs to the end of the stream.
• Error::Unsupported for recognized formats that can't be decoded, such
  as Opus without the opus feature (the message names the feature).
• Opus decoding (Ogg and WebM) behind the optional opus feature, using the
  pure-Rust ruopus decoder. Its output matches libopus to within float
  rounding (121 dB SNR on speech). It is off by default because it adds
  about 20 crates and needs Rust 1.95.
• cargo bench --bench throughput, a process_file example, and
  cargo fuzz targets for decoding and the pipeline.

Changed

• Breaking: AudioPipelineCoordinator is now Pipeline. Its methods
  take &mut self and return the cleaned ProcessedChunks instead of only
  a count. process_frame is now process_bytes.
• Breaking: VadDetector::new takes only a VadConfig.
• Breaking: VADStats is now vad::VadStats.
• Breaking: the VAD decides on frame energy relative to a tracked noise
  floor and nothing else. The spectral-flux score, combined threshold and
  their settings never changed a decision once the relative energy gate was
  in place, so they are gone: VadConfig loses energy_smoothing
  (now noise_smoothing), flux_smoothing, flux_floor,
  threshold_smoothing, activation_margin, release_margin,
  base_threshold, energy_weight, flux_weight and fft_size(), and
  gains activation_ratio (default 2.5) and release_ratio (default 1.3).
  The hysteresis between them raises F1 on real speech from 0.946 to 0.958
  and splits speech in heavy noise into fewer pieces.
  VadMetricsSnapshot::adaptive and AdaptiveThresholdSnapshot are
  replaced by VadMetricsSnapshot::noise_floor.
• SpeechChunk::confidence is now how far the segment's speech rises above
  the noise floor (0 at the floor, 1 at 20 dB above). It used to be 1.0 for
  almost every segment.
• Breaking: ConversionMetadata reports a single conversion_time and
  original_channels is a u16.
• Breaking: AudioFormat names the codec, not the container, and is
  #[non_exhaustive]: WebM is gone (a WebM file reports Vorbis or
  Opus), Alac and Adpcm are new. FormatDetector::detect always
  probes the stream and fills in all metadata;
  FormatDetector::detect_with_metadata, AudioMetadata::format_only, and
  AudioMetadata::with_properties
  are removed.
• Decoding no longer copies the input buffer.
• Breaking: DcHighPassFilter::process takes Option<VadContext> by
  value, like NoiseReducer::reduce.
• Breaking: configuration validation errors are Error::Configuration
  everywhere; Error is #[non_exhaustive].
• The VAD noise floor is seeded from the quietest frame of the first
  300 ms and keeps adapting during speech. Before, a recording that started
  with noise above the energy gate was classified as one long speech
  segment.
• The default hangover_frames is 20 (200 ms) instead of 3 (30 ms; the
  docs said 60 ms), and hangover frames no longer extend a segment's end.
  Segments get 50 ms of padding on each side (VadConfig::speech_padding,
  new). On real speech in three languages, scored against Silero VAD and a
  human annotation, a sentence is no longer split at every gap between
  words (e.g. 6 segments instead of 56 on a 30 s meeting clip) and F1 goes
  from 0.89 to 0.95.
• VadStats::speech_frames counts frames that measured as speech; hangover
  frames now count as silence.
• Because the noise floor now adapts, the default VAD energy gate is 40x
  lower (0.0005 instead of 0.02). The old value missed 60% of speech
  peaking at -6 dBFS and all speech at -20 dBFS in a synthesized test set.
• Channels are mixed to mono before resampling instead of after.
• Resampling uses a Kaiser-windowed sinc filter instead of linear
  interpolation. A 12 kHz tone at 48 kHz used to alias to 4 kHz at full
  level; it is now attenuated by about 100 dB.
• Licensing: the crate is now MIT only (0.1.x was MIT OR Apache-2.0).
• Edition 2024. Cargo now resolves dependencies against rust-version, so
  1.87 builds without pinning anything.
• Library code no longer logs at info/warn on every call or warns when a
  stage takes longer than a fixed wall-clock budget.

Fixed

• Resampler response documentation now distinguishes downsampling, lower-rate
  input filtering, and unfiltered 16 kHz passthrough; the codec support table
  no longer claims mu-law or A-law decoding in CAF.
• The noise floor's slow rise during speech now has the same time constant
  (5 s) at any frame hop. It was a fixed per-frame factor, so with
  frame_overlap: 0.75 steady noise under speech was absorbed in half the
  time, and with no overlap in twice the time. Defaults are unchanged.
• VadConfig::validate rejects an infinite min_speech_energy, which passed
  the sign check and silently turned detection off.
• PreprocessingConfig::validate rejects a non-finite high-pass cutoff or
  dc_bias_alpha, which previously passed and turned the filter output into
  NaN.
• A ChunkerConfig built as a struct literal with a sub-millisecond
  target_duration is now Error::Configuration from Chunker::chunk
  instead of a divide-by-zero panic.
• NoiseReducer now carries the last frame of the stream as real left-hand
  context and keeps its STFT frames on one hop grid across calls, so a
  continuous signal split across calls (at any sample, in blocks of any size)
  no longer has a discontinuity at the leading edge of each chunk once a noise
  profile is learned (previously each call zero-padded its own left edge and
  restarted its frame grid). The trailing
  edge of a call still lacks future context and is re-covered on the next
  call; eliminating that residual entirely would require an output-lagging
  (buffered) API, which is a separate change.
• A truncated stream (cut mid-packet) is now rejected with InvalidInput
  instead of returning the partial audio decoded so far as Ok.
• Symphonia 0.6.1 or newer is required so a malformed metadata length cannot
  trigger an allocation for a payload that is absent from the input.
• The decoder caps how many samples one decode may buffer (64 Mi interleaved,
  256 MiB), and how many 16 kHz samples it will resample to, so a
  decompression bomb — a tiny compressed file declaring hours of audio — is
  refused instead of allocating gigabytes.
• The pipeline no longer emits the start of a speech segment that is still
  in progress as a silence chunk.
• Pipeline speech chunks emitted from a window with no silence of its own
  (long speech held past max_speech_hold, or the end of a stream) had no
  snr_db. They now use the level of the last silence in the stream.
• NoiseReducer no longer fades each chunk in from near zero at its first
  samples (about -23 dB on a steady tone with 500 ms chunks).
• The fourth-order high-pass filter now has Butterworth stage Q values; the
  cutoff is at -3 dB instead of -6 dB.
• VadStats::speech_frames counted segments instead of frames, and
  silence_frames was always zero.
• QualityMetrics::spectral_centroid is now computed from the spectrum. It
  used to be a weighted time index, not a frequency.
• VadDetector::reset now also resets the noise floor.
• VadDetector::detect keeps samples that don't fill a frame for the next
  call. Before, each call started a new frame grid, so blocks under about 64
  samples never detected speech.
• With frame_overlap above 0.5, consecutive speech segments could overlap
  and the chunker emitted the overlapping audio twice. Segments are now
  disjoint, and Chunker merges overlapping input segments.
• Decoding no longer returns empty or partial audio without an error for
  chained Ogg streams, or for files where many packets fail to decode.
  Same-layout chained streams decode fully; a sample rate or channel count
  change is Error::Unsupported.
• A panic inside Symphonia on malformed input (found by fuzzing: a
  Matroska element with a zero length byte) is returned as
  Error::InvalidInput instead of unwinding into the caller.
• StandardAudio samples are now guaranteed finite and within [-1, 1]:
  NaN or infinite input is an error, and louder-than-full-scale float input
  or resampler overshoot is clamped.
• Inputs declaring a sample rate outside 4-768 kHz are rejected with
  Error::Unsupported. Fuzzing found a 10 KB file claiming 1.04 GHz that
  made the resampler build a 1.7 GB filter and run for 40 s.
• NoiseReductionConfig::validate rejects a hop shorter than one sample,
  which made reduce loop forever.
• Symphonia 0.6 fixes an untrusted allocation in the MP4 reader: a 5 KB
  M4A with a corrupt sample table could make 0.5 try to allocate 32 GB and
  abort the process (found by fuzzing).

Removed

• Breaking: buffer::AudioBuffer, types::AudioChunk, the fixtures
  feature and fixtures module, Chunker::default() as an inherent method
  (use Default), ProcessedChunk::samples_without_overlap,
  ProcessedChunk::overlap_next (it duplicated the next chunk's
  overlap_prev), ProcessedChunk::total_samples_with_overlap,
  ConversionMetadata::has_performance_issue,
  AudioFormat::is_container_format, ProcessingResult::backpressure_active,
  and StageLatencies::broadcasting_avg.
• glob, serde_json, parking_lot, and rand dependencies; hound is
  now dev-only.