PACC-T: PARALLEL ACOUSTIC CONFOUND CORPUS, TELECOMS
EDGE CASES

Things a user should know that the per-clip params do not make obvious.
Figures are over the 5,992 released source clips.

1. EVS DTX variants coincide on clips without pauses
   DTX only changes frames the encoder classes as inactive. On clips
   with no such frames, the variants produce identical audio:

     rate      all three identical   fixed == adaptive   fixed != off
     9.6k             1,129               2,550              4,863
     24.4k            1,125               2,552              4,867
     SWB 48k          1,043               2,504              4,949

   The three evs_24400_dtxfixed tandems equal their evs_24400_nodtx
   counterparts on the same 1,125 clips. Compare DTX variants on the
   clips where they differ, or weight results accordingly.

2. EVS dtxfixed, version note
   In v2.2 the three evs_*_dtxfixed conditions and the three
   evs_24400_dtxfixed tandems were re-encoded with -dtx 8. Earlier,
   unreleased builds held DTX-off audio under that label. Every v2.2 row
   in _meta/params records EVS_cod's own confirmation:
   "DTX: ON, CNG update interval = 8 frames".

3. LC3 low-level pass-through
   For 2 clips, lc3 output is bit-identical to resample_ctrl_16k:
     AMI  EN2006b_C_0031_585.87_588.52
     AMI  ES2007d_D_0039_1117.70_1120.36
   Both are quiet (P.56 speech level -54.7 and -50.0 dBFS) but within the
   retained range.

4. Excluded sources
   8 AMI source clips with no usable speech are excluded from all of
   PACC (both PACC-T and PACC-P). See excluded_sources.csv.

5. Inherited source clipping
   350 AMI source clips contain full-scale samples, clipped in the
   original recordings. They are encoded as they are, not repaired.

6. Options with no effect (recorded as run)
   - codec2_1300: the recorded arguments include -b:a 700, which libcodec2
     ignores; the encoder ran in its default 1300 bps mode, hence the name.
     codec2_700 uses -mode 700C.
   - speex_8k "-vbr off" and aac_* "-vbr 0" are not consumed by those
     encoders; both ran at default rate control.
   The params keep the arguments exactly as run, so re-running them
   reproduces the files.

7. Decoders
   Only ilbc pins its decoder (-c:a libilbc). All other conditions use
   ffmpeg 8.1's default decoder for the stream, as recorded.

8. How the params were obtained
   - 6 conditions (the v2.2 re-encodes) were logged per clip at encode
     time.
   - The other 45 were recovered: the recorded command for each condition
     was re-executed on 60 clips (20 per source pool) and reproduced the
     stored audio bit for bit, 2,700 of 2,700; per-clip rows were then
     filled from the files (sha256, sample rate, frames).
   - The column params_origin (logged / recovered) states which applies
     to each row; verified_by names the verification run.
   - v1 tandems read their first hop as WAV before the FLAC conversion;
     the params give the FLAC path, which holds identical PCM.