spokenform

spokenform is a single-language written-to-spoken text normalizer for speech applications. It expands abbreviations and units, verbalizes structured numeric forms, protects literal ranges, and records stage-by-stage provenance and offset maps.

Documentation

Installation
- Runtime
- Optional spaCy support
- Documentation
- Optional Lexhint integration
Languages
- Complete runtime registry
Language support matrix
- German scope
- Swedish scope
- Vietnamese scope
- Thai scope
- Russian scope
- Identifier rules
- Ownership and safety
Generated language coverage
Quickstart
- Japanese, Korean, and Chinese
- Swedish
- Vietnamese
- Thai
- Configuration object
Architecture
- Ownership boundary
- Structured precedence
- Recognition policy boundary
- Optional Lexhint boundary
- Semantic segment boundaries
API reference
- Preparation
- Language identifiers and number backends
- Configuration policy modes
- Result models
- Annotation adapters
- Number normalization
- Interpretation policy
- Lexhint evidence provider
Speech profiles
- Basic usage
- Isolation and precedence
- Current boundaries
kokorog2p migration boundary
- Ownership audit
- Preferred adapter surface
- Compatibility cleanup
PiperG2P integration boundary
- Prepared text
- Number ownership
- Raw phoneme blocks
- Piper overrides and coordinates
- Annotations
- CI contract
- Dependency direction
- Non-goals
Command-line interface
- Interpretation and domain controls
spaCy integration
- Current effect with the released abbr2words structured API
- Load an installed model by name
- Inject an application-owned pipeline
- Supply annotations directly
- Required token contract
- Error behavior
Protected text
Offset mapping and provenance
Mapping and provenance
Examples
- Interactive Jupyter notebook
Limitations
Limitations and readiness gates
- Recognition modes and specialist domains
Benchmarks
- PolyNorm diagnostic benchmark
- Proteno benchmark
- Google TN benchmark
- Async Voice TTS Normalization Benchmark
- Spokenform Gold
- Lexhint A/B comparison
Release checklist
- Spokenform Gold benchmark consumption
Changelog
- [v0.4.5] - 2026-09-21
- [v0.4.4] - 2026-09-15
- [0.4.3] - 2026-09-14
- [v0.4.2] - 2026-09-13
- [v0.4.1] - 2026-09-13
- [0.4.0] - 2026-09-13
- [0.3.7] - 2026-09-12
- [0.3.6] - 2026-09-01
- [0.3.5] - 2026-08-31
- [0.3.4] - 2026-08-30
- [0.3.3] - 2026-08-27
- [0.3.2] - 2026-08-26
- [0.3.1] - 2026-08-23
- [0.3.0] - 2026-08-19
- [0.2.8] - 2026-08-18
- [0.2.7] - 2026-08-16
- [0.2.6] - 2026-08-14
- [0.2.5] - 2026-08-12
- [0.2.4] - 2026-08-11
- [0.2.3] - 2026-08-10
- [0.2.2] - 2026-08-09
- [0.2.1] - 2026-08-07
- [0.2.0] - 2026-08-07
- [0.1.0] - 2026-08-07

Historical

spokenform 0.2.3 downstream handoff

Scope

The package accepts plain text and one selected processing language. Language detection, mixed-language segmentation, SSMD or other markup parsing, and phoneme generation are intentionally outside the package.