ASR Bias Builder

Build ASR bias artifacts from presentation decks for Whisper and Google Speech-to-Text

ASR Bias Builder extracts high-value entities from PDF/PPTX decks and produces:

• deck_terms.txt – Whisper initial_prompt hotwords
• phrase_set.json – Google Speech-to-Text v2 Speech Adaptation PhraseSet
• Structured LLM candidates + verification telemetry for auditability

Quick Start

pip install -e .
asr-bias-builder pipeline deck.pdf

Artifacts land in ./asr-bias-output/<deck-name>/ by default (deck text, seeds, verified terms, prompt list, phrase set, review markdown) and the cross-run summary lives at ./asr-bias-summary.csv. Override the locations any time via --output-dir / --summary-csv.

Features

• Deterministic PDF/PPTX extraction with OCR normalization
• Seed mining with configurable filters, section weighting, and per-deck overrides
• LLM integration via Claude Code CLI (stdin, streaming JSON, read tool)
• Verification layer merges deterministic + LLM terms with alias tracking
• Artifact builders for Whisper prompts and Google Speech Adaptation PhraseSets
• Review markdown + CSV summaries for production runs
• Docker, CI/CD, docs, tests, and examples ready for GitHub publishing

Repository Layout

asr-bias-builder/
├── asr_bias_builder/        # Python package
├── config/                  # Default + example configs
├── docs/                    # User + architecture docs
├── examples/                # Synthetic sample decks & configs
├── scripts/                 # Shell helpers (make_bias.sh, claudecode, etc.)
├── tests/                   # Pytest suite and fixtures
├── docker/                  # Dockerfiles & compose for dev/prod
└── .github/                 # Workflows, templates, Dependabot

See the docs for detailed installation, configuration, and architecture notes.