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.

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 stock SphinxTrain. On SLT-55, aggregate WER and error count are identical even though 26 of 55 per-utterance error rows differ. 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.

What pstrain does

pstrain provides an in-process 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.

pstrain supports macOS and Linux, and now Windows. See support policy for details.

Quickstart

From PyPI

pip install pstrain

Train with your own audio, prompt list, pronunciation dictionary, phoneset, and filler dictionary:

pstrain train /tmp/pstrain-project \
  --audio your-audio/ \
  --prompts your-prompts.txt \
  --dictionary your-dictionary.dict \
  --phoneset your-phoneset.txt \
  --filler-dict your-filler.dict \
  -j 1

Prompt lists are assumed to be pre-normalized to match the lexicon.

From a checkout

Clone the repository and install it in editable mode with the decoding extra:

git clone https://github.com/lenzo-ka/pstrain.git
cd pstrain
python -m pip install -e ".[test]"

The checkout includes a small CMU ARCTIC fixture for a complete local run. Set up a project and train the default eight-Gaussian context-dependent model:

pstrain train /tmp/pstrain-demo \
  --audio tests/fixtures/mini_arctic/wav \
  --prompts tests/fixtures/mini_arctic/transcription.txt \
  --dictionary tests/fixtures/mini_arctic/dictionary.dict \
  --phoneset tests/fixtures/mini_arctic/phoneset.txt \
  --filler-dict tests/fixtures/mini_arctic/filler.dict \
  -j 1

The command stores separate typed training and decoder transcripts. Decode the held-out set without a language model directly—no transcript conversion is needed:

pstrain test cd-8g --project-dir /tmp/pstrain-demo --no-lm

For a project of your own, pstrain train --help describes the accepted audio, prompt, dictionary, phoneset, and configuration inputs. The lower-level setup, validate, and build commands remain available for decomposed workflows. The getting-started guide continues from project setup.

The default pstrain train target is cd-8g. Use --target ci-1g when only the faster context-independent bootstrap model is wanted.

Benchmark pin

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

Documentation

• Getting started
• 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.