[bugtoo]

Summary

Adds two output controls to extractaudio:

• -r 8000|16000: output sample rate. At 16000 Hz, G.722 legs keep their wideband audio.
• -b KBPS (Opus) and -C LEVEL (any libsndfile format): encoder bitrate or compression level.

Also fixes two small bugs found along the way. Defaults are unchanged: without the new options the output is byte-identical to the current code.

Why

We convert PBX call recordings with extractaudio -s so that caller and callee land on separate channels. Each party can then be transcribed separately and the transcript split by speaker. Two limits get in the way:

• 8 kHz only. G.722 legs are decoded down to narrowband, so wideband calls lose half their bandwidth before speech recognition.
• No bitrate control. With -F ogg -D opus, libsndfile picks its own default (about 59 kbit/s for stereo). At low bitrates Opus merges the two channels into one, which defeats the stereo split, so we need to choose a rate that keeps them separate at a reasonable size.

Implementation notes

• -r 16000:
  • Timing stays in 8 kHz RTP ticks; each tick produces rate / 8000 output samples, and the channel start offset is computed at the output rate.
  • G.722 is decoded at its native rate.
  • Narrowband legs (G.711, G.729, GSM) are upn-windowed half-band interpolator. It is flatto 3 kHz, -1.6 dB at 3.4 kHz, with a fixed 1 ms delay.
  • Upsampling narrowband legs lets one stereo file combine legs that use different codecs.
• -b / -C:
  • Both set SFC_SET_COMPRESSION_LEVEL after sf_open() and before the first write.
  • For Opus, -b converts kbit/s to the level using libsndfile's mapping: ((1 - q) * 250000 + 6000) bit/s per channel.
  • -b requires -D opus; -b and -C are
  • Option errors use warnx() and the common exit path, like the existing options.
• Fixes:
  • generate_silence() advanced the G.729 ouer 80-sample frame.
  • The "written" count printed for mono output was halved.

Testing

New tests

• tests/extractaudio/badargs: invalid -r (11025, 16000x), -b outside 6–512 or without -D opus, -b combined with -C, and -C 2.
• tests/extractaudio/extractaudio16k (new, r runs -r 16000 in mono and stereo on everyexisting recording and checks 16 SHA-256 checksums plus the RTP stats output.

How we ran them. The environment mirrors 24.04 with bcg729 1.1.1, libg722, libsrtp2_x_devand libsndfile 1.2.2 built from source, pluslibssl-dev`, which the GitHub runners have preinstalled.

./configure --enable-basic-tests && make
make -C tests check TESTS="extractaudio/baio1 extractaudio/scanextractaudio/extractaudio16k"

• Results: all four suites pass with gcc and with clang, and the changed files build without warnings.
• Defaults unchanged: the existing extractaudio1 checksums still pass.
• Deterministic output: the 16 kHz checksums came out identical in two separate clean builds.

Opus. Not covered by the checksum tests: scripts/do-build.sh builds libsndfile without installing Opus or Ogg libraries first, so CI likely has no Opus encoder. It was checked by hand on real two-leg recordings, on Linux and on FreeBSD 14.4, using libsndfile 1.2.2 built with Opus.

• 16 kHz conversion:

  • exactly twice the frames;
  • narrowband samples pass through unchanged;
  • energy above 4.2 kHz is 57–77 dB below the speech band.

• Stereo channel separation by bitrate:

  Option	Separation
  -b 16	0 dB (Opus downmixes to mono)
  -b 24	24 dB
  -b 32	48 dB

  We use -s -r 16000 -b 32 -F ogg -D opus.

[sobomax]

@bugtoo thanks for the PR — merged into master. Before merging I made some changes on top of your patches and squashed them into your two commits, so the history still shows you as author. Here's what changed:

Decoder

• Upsampler alignment. The upsampler delayed narrowband audio by 1 ms, while G.722 at 16 kHz had no delay. That put a G.722 leg and a narrowband leg 1 ms apart in stereo. Its history also carried across codec switches within a leg, so a G.722 → PCMU switch played back old samples and dropped the last 1 ms. It now has no delay and restarts on each switch, so N samples at 8 kHz always become exactly 2N samples at 16 kHz.
• 8 kHz regression. Enlarging obuf for 16 kHz also raised the per-packet decode limit at 8 kHz. Oversized garbage packets marked as G.722 were therefore decoded further than before, which changed the rtp_analyze/sipit31 checksum. Decoding is now capped at DECODER_MAX_TICKS per packet at either rate, so 8 kHz output is byte-identical to before.
• Buffer sizes now come from DECODER_MAX_TICKS and DECODER_MAX_RMULT. The old literal 4000-tick cap on inserted silence is now DECODER_MAX_GAP, with a compile-time check that it fits the buffers.

Options

• -b is checked against libsndfile's Opus range of 6–256 kbps per channel. Previously a value outside it was silently clamped, e.g. mono -b 512 gave 256 kbps.
• -b and -C reject nan and empty values. strtod() accepts them, and the option was then silently ignored.
• -r 16000 rejects 8 kHz-only data formats (GSM 06.10, G.721, G.723). Without -D, mono output at 16 kHz defaults to MS ADPCM instead of GSM 06.10.
• -C on a data format without compression levels is now an error instead of a warning, and the half-written output file is removed.

Tests

• extractaudio1: extractaudio16k is merged into it. The 16 kHz run now also covers raw, w64 and -n. It also checks that for every narrowband recording, each channel of the 16 kHz output, taken every other sample, matches the 8 kHz output exactly.
• mixed_codec: a new recording spliced from existing ones. One leg switches G.722 → PCMA → G.722 within one SSRC, and the other leg is PCMU, so it covers codec switches and mixed-rate stereo.
• encopts: a new test checking that -b gives the requested Opus bitrate within 25% at both rates, mono and stereo, and that a higher -C gives a smaller file. It requires a libsndfile with Opus support.
• badargs: the -r/-b/-C cases now run against a real recording. With /dev/null as input they would have failed even if the bad value had been accepted. Each case must also leave no output file behind.

CI

• The test jobs build libsndfile from source. It previously built without Ogg/Vorbis/FLAC/Opus because their development packages weren't installed. The packages are now installed, and the build stops if libsndfile still comes out without them.

[bugtoo]

Thank you @sobomax for fixing my PR and merging it!