Historical, incomplete evidence: M4b SLT runs
Corpus: the fixed 1,043-utterance training split in
pstrain-cmu-arctic-slt. The source log was
/Volumes/experiments/pstrain-parity/pstrain-cmu-arctic-slt/training.log.
Arm comparability: incomplete. The recorded arms used the same stated corpus and CD-untied stage, so the artifact-size and graph-growth observations below are paired on that limited basis. The baseline raw log was not retained, however, so its complete run identity, configuration, skip identities, and resource provenance cannot be reconstructed independently. Treat these results as historical implementation evidence, not as a reproducible benchmark or a controlled performance comparison.
Skip counts
Run |
Result |
|---|---|
terminal-filler fix only (baseline) |
224 / 1,043 skipped at CD-untied iteration 3 |
two-sided graph only |
198 / 1,043 skipped |
graph plus dictionary-domain initialization |
10 failed attempts / 1,043 (five utterances, original plus widened retry) |
The five utterance identities and the 1e-100/1e-200 direct-probe
results are in m4b-slt-skips.tsv. The normal 1e-90 run automatically
retries once at 1e-100, hence ten failed attempts but five skipped
utterances. All five also fail with the effectively-off 1e-200 beam.
The baseline raw log was not retained as a standalone lane artifact, so the 224 identities cannot be reconstructed unambiguously from the shared log, which contains interleaved output from several forced runs. This file records that evidence limitation rather than manufacturing an identity list.
Reproduction commands
PSTRAIN_LIB_PATH=/path/to/libpstrainc.dylib \
pstrain build cd-untied --project /path/to/pstrain-cmu-arctic-slt \
--config parity --force
# Direct stability probes use the saved cd-untied model and each utterance
# listed in m4b-slt-skips.tsv with BWConfig(a_beam=BEAM, topn=1,
# multipron=True), for BEAM in 1e-90, 1e-100, and 1e-200.
Growth
Metric |
Before |
After |
Change |
|---|---|---|---|
Untied triphone rows |
9,786 |
10,052 |
+266 (+2.72%) |
mdef |
491,359 B |
504,926 B |
+13,567 B |
means |
4,598,636 B |
4,723,124 B |
+124,488 B |
variances |
4,598,636 B |
4,723,124 B |
+124,488 B |
mixture weights |
117,976 B |
121,168 B |
+3,192 B |
transition matrices |
1,984 B |
1,984 B |
0 |
parameter files total |
9,309,224 B |
9,561,400 B |
+252,176 B (+2.71%) |
CD-untied wall time |
2.4 s |
2.2 s |
-0.2 s (run noise) |
Peak RSS was not captured. At the engineered two-variant boundary, the one-sided graph had 12 phone slots/14 directed edges and the two-sided graph has 14/16: two local additions across two ambiguous positions.