pstrain Documentation

pstrain is an acoustic model training toolkit.

User Guide

• Getting started
  • Install the current checkout
  • Train a small model
  • Install a published release instead
  • Further workflows
• Input formats
  • Pronunciation dictionary
  • Transcripts
  • Phoneset
  • Audio
• Examples
  • Project Setup
  • Project Validation
  • Model Creation
  • Data Structures
  • Log-domain Math
• Alignment mass and coverage report
  • Where it appears
  • Outcomes
  • What is counted
  • Flags
  • Size
• Support and dependency policy
  • Platforms
  • Dependencies
• Package replacement safety
  • Concurrent pathname changes
• Glossary
  • Model Types
  • Core Concepts
  • Features
  • Training Algorithms
  • Model Parameters
  • Files and Formats
  • Training Stages
  • Decision Trees
  • Abbreviations Reference

API Reference

• API Reference
  • Library API
  • Alignment API
  • Checkpoint recovery API
  • Configuration API
  • Corpus API
  • Diagnostics API
  • Dictionary API
  • Language model API
  • One-command Training API
  • Packaging API
  • Pipeline API
  • Training Steps API
  • Testing API
  • Timings API
  • Tutorial API
  • CLI API
  • Data Utilities
  • Configuration Reference

Design Notes

• Design documentation
  • Current architecture
  • Domain references
  • Past decisions
  • Decisions reversed (kept here for context)
• Parity and declared deviations
  • Declared register
  • Inventory under multiple pronunciations
  • Training front-end boundary
  • Alignment recovery and final silence
  • Checkpoints, sharding, and reduction
  • Dropped base pronunciations in the alignment dictionary
• Canonical configuration
  • Decision summary
  • Current state
  • Truthfulness invariant
  • Proposed model
  • Compatibility and C2 landing order
  • Open decisions for Kevin
• One-command training
  • Goal
  • Input contract
  • Pre-normalized prompts are the default
  • Handoff to setup and build
  • Failure UX
  • CLI verb choice
  • Compatibility and rollout
  • Open decisions for Kevin
• Pipeline runner
  • What it is
  • How it works
  • Why we built our own
  • What we explicitly don’t support
  • Multi-pronunciation training
  • Stage-specific Baum-Welch control
  • Adding a new pipeline node
  • Testing
  • Experimental split variance regularization
  • Explicit checkpoint recovery
• Training Pipeline Plan
  • Implementation Status Grid
  • Key Insight: CFFI Bindings Are Complete
  • What’s Missing: Step Orchestration
  • Architecture
  • What’s Complete ✅
  • What Needs Implementation ⏳
  • Implementation Order
  • File Locations
  • Success Criteria
• Multi-pronunciation training
  • Status
  • What it does in one paragraph
  • Why the prior trainer was wrong
  • How it works
  • As-built layout
  • Pipeline graph shape
  • Operations
  • What we don’t do
  • M4b SLT growth measurement
  • Empirical signal
  • Commit history
  • Relationship to the upstream graph builders
  • Future work
• Baum-Welch normalization policy
• Baum-Welch sharding contract
  • Generated gate scope
  • Hand-written limitations and interpretation
  • Cross-count characterization
  • Multipron activation and worker lifetime
• Failed-alignment policy
  • Deliberate vendored divergence
• Optional final transcript silence
  • Deliberate vendored divergence
  • Graph ownership
  • What mandatory final silence can measure
  • Scope
• CLI-to-library boundary
  • What these checks are, and are not
  • Runtime observation
  • Static addition
  • Combined coverage and limits
• Native boundary: what is contained, and what is not
  • Phase contract: contained-all-operations
  • Complete-model value validation
  • Exception hierarchy
  • Lifecycle
  • No silent fallback
• Numerical-correctness harness
  • Purpose
  • Golden trajectory
  • Floating-point contraction checks
  • Five choke points
  • Acceptance debts
  • Per-pass utterance exclusions
  • Reproducibility
• Decision-tree and tied-state comparison
• Per-step profiling
• Source-determinable re-roll dispositions
• Project Setup Design
  • Overview
  • Project Structure
  • Setup Requirements
  • Setup Process
  • CLI Command
  • Implementation for pstrain
  • Configuration File Format
  • Validation
  • File Format Details
  • Integration with Workflows
  • Next Steps After Setup
  • Design Notes
  • Testing with CMU Arctic
• CI Model Building Requirements
  • Overview
  • Prerequisites (Dependencies)
  • CI Training Process
  • CI Training Parameters
  • Workflow Dependencies
  • Implementation for pstrain
  • Key Programs Needed
  • Output Products
  • Implementation Notes
  • Next Steps
• Testing with CMU Arctic Corpus
  • Overview
  • CMU Arctic Structure
  • Process for Testing
  • Implementation Notes
  • Testing Workflow
  • Required Files Summary
  • Notes
• Terminology
  • Transcription vs Alignment
  • Usage

Benchmarks and Evidence

• Arctic benchmark pin
  • Measurement identity
  • Pin conditions
  • Provenance correction
  • Comparability
  • Baseline
  • Forward gate
  • Decode-path transport
  • Training skips and decode coverage
  • Condition contract maintenance
  • Replicability
• Arctic oracle provenance and reconstruction
  • Keeping the live comparison resource-matched
  • Procedure for an independently derived stock arm
• Historical, incomplete evidence: M4b SLT runs
  • Skip counts
  • Reproduction commands
  • Growth

Contributing

• Development
  • Canonical verification command
  • Individual build and test commands
  • Changing the native interface
  • Building documentation
  • Code Quality
• Detached runs

Indices and tables

• Index

• Module Index

• Search Page