pstrain — Peace Train

The name pstrain is short for PocketSphinx train; it is pronounced "peace train."

pstrain is a toolkit for training continuous-density HMM/GMM acoustic models. It combines a C library for the numerical work with a Python interface for project setup, configuration, and training pipelines.

pstrain descends from CMU SphinxTrain and the wider CMU Sphinx lineage. We are grateful to the developers, researchers, institutions, and community members whose work made this project possible.

What pstrain does

pstrain provides a training pipeline for:

• acoustic feature extraction and corpus splitting;
• flat initialization and Baum–Welch training;
• context-independent and context-dependent models;
• Gaussian splitting, decision-tree state tying, and model packaging;
• forced alignment and PocketSphinx decoding with WER/CER evaluation; and
• multiple-pronunciation training, where pronunciation variants participate as parallel paths in each utterance's training graph.

Quickstart

Start with the terminal walkthrough: it covers Python and compiler prerequisites, an isolated installation, and a small training run using the included audio, followed by decoding and packaging. No Python programming or corpus download is needed for that example.

Training requires macOS or Linux and Python 3.11+. Native Windows supports the package and native tools, but cannot run the training pipeline; see the support policy.

The guide distinguishes building this checkout from installing a published release, and starts with the faster ci-1g model before the default cd-8g.

Benchmark pin

The ARCTIC benchmark pin freezes the corpus, training modes, model identities, decoder, and per-utterance measurements used as the comparison baseline.

pstrain rebuilds SphinxTrain with its C core vendored and its orchestration reimplemented. On the shared CMU Arctic benchmark, pstrain under its shipped defaults except for split.test_count=0 (including multiple-pronunciation training) shows no statistically significant regression on either cell versus the preserved upstream oracle, historically attributed to stock SphinxTrain; see its provenance limits. The arms share one matched decode path but are NOT COMPARABLE for implementation attribution because the oracle's producing lineage is unknown. The two cells carry different weight. The 3,395-utterance set is cross-speaker, so it measures what the model does on voices it was not trained on: there pstrain records 22,703 errors (75.7221% WER) against 22,638 (75.5053%) for that oracle — behind by 0.2168 percentage points, a margin the data cannot separate from zero (95% paired interval [-0.2414, +0.6713]). SLT-55 is a same-speaker held-out cell of 55 utterances; there pstrain is ahead, 139 errors (27.0955% WER) against 144 (28.0702%), with 23 of 55 per-utterance error rows differing, but its interval of [-3.6965, +1.6575] is consistent with anything from a 3.7-point gain to a 1.7-point regression and so settles little on its own. The larger cross-speaker set is where the evidence is. The benchmark pin and its machine-readable record preserve the evidence and measurement conditions. Run make verified for the repository's aggregate verification command; its Arctic gate authenticates the pinned measurement conditions against the current checkout and re-derives the recorded paired statistics from the stored per-utterance rows without re-decoding. The end-to-end pstrain-versus-baseline WER comparison is produced by the full scripts/bench_arctic.py benchmark run, not by make verified.

Documentation

• Getting started
• Package replacement safety
• Input formats
• Examples
• API reference
• Design documents
• Development guide
• Support and dependency policy

The documentation can be built locally with make -C docs html.

License

New pstrain code is available under the BSD 2-Clause license. CMU-derived code retains its applicable CMU BSD-style license alongside the license for pstrain modifications. See LICENSE, csrc/LICENSE.sphinx, and THIRD_PARTY_NOTICES.md for the complete terms and credits.