Changelog

Added a versioned request API compatibility contract and installed-wheel checks

Added transcript-guided English voice enrollment with reusable, model-bound reference-voice conditioning

Added transcript-free Inno voice enrollment that exports reusable Kokoro voice packs

Changed voice blending to support canonical = syntax, multi-voice linear interpolation, and row-wise SLERP fallback

Raised PyKokoro’s minimum supported OnnxVoice version to 0.2.0 while keeping its model-local voice API unchanged

Fixed forwarding of SynthesisConfig.cache_dir to OnnxVoice for managed model and voice assets

Fixed the long-text example to consume its result iterator before validating the single rendered request

BREAKING: Removed the no-op EspeakConfig and espeak_config options

Documented cache, discovery, asset progress, and request controls with expanded API guides and grouped examples

Improved import ordering in short-sentence benchmark scripts

Added opt-in token-safe splitting for oversized requests with one rendered result

BREAKING: Changed PyKokoro to render prepared requests as independent results

Changed model and ONNX session management to use OnnxVoice

Changed and validated the supported companion-library minimum versions

Documented the request migration, long-text controls, and retained compatibility aliases

Improved regression and integration coverage for the engine boundary

Added output loudness measurement with pipeline integration and configurable normalization

Added configurable voice loudness measurement benchmark with policy-driven corpus evaluation

Added voice level calibration data and per-voice loudness normalization

Documented public API type hints and expanded voice documentation in README

Added an English 7M Distill example and preserved asset progress callbacks during pipeline preparation

Added metadata-only named-lexicon discovery and made German espeak the implicit lexicon default

Added localized short-sentence catalogs, punctuation-aware forms, German Crane readiness, and catalog diagnostics

Changed short-sentence rendering to use timestamp geometry, legal boundaries, adaptive cuts, and controlled fallbacks

Improved short-sentence quality tooling with reproducible latency, diagnostics, parameter sweeps, and reports

Added explicit runtime support for the Vietnamese Ngoc Huyen model profile

Added configurable automatic pauses at detected parenthetical boundaries with G2P propagation and example coverage

Added an explicitly selectable English 7M distill model profile without changing automatic voice selection

Changed the Vietnamese An Phunl profile to runtime-ready status with its published default voice

Fixed the default voice selected for the Vietnamese ContextBox profile

Fixed the all-voices showcase to handle voices with incomplete registry metadata using language-aware fallback details

Fixed Chinese v1.1 vocabulary handling across tokenization and ONNX audio input generation

Added automatic high-confidence clausal-comma pause detection for dependency-aware English segmentation

Added configurable short automatic pauses around Phrasplit-detected parenthetical asides

Documented clausal-comma pauses with a runnable English A/B example and local spaCy setup instructions

Added GitHub-backed Russian model profiles with voice-aware model selection and compatibility validation

Changed pipeline-owned stages to follow model variant changes while preserving custom stages after replacement failures

Changed short-sentence defaults to use wrap mode without timestamps and warn on explicit phrase-mode fallback

Changed the kokorog2p compatibility floor to 0.9.9

Fixed traced exceptions being replaced while still recording timing events and debug logs

Added observable lazy runtime asset downloads with byte-level progress callbacks and console reporting

Added per-voice metadata and an all-voices showcase for dynamic inventory and labeled audio generation

Added standard Python logging for backend, asset, synthesis-stage, and runtime lifecycle diagnostics

Added runtime-ready European Portuguese model support with regional pt-PT routing and asset provisioning

Changed release manifest resolution to select model assets by quality and validate package dependency versions

Added runtime-ready German Anna model support with automatic discovery, asset resolution, and example coverage

Added automatic language-aware routing for mixed-language documents

Added explicit G2P fallback and named-lexicon controls

Improved mixed-language frontend selection and caching

Added public resolve_pipeline_config for metadata-only automatic model, source, quality, and voice selection

Added configurable inference caching, traceable ONNX performance metrics, and a German CPU benchmark

Added metadata-only model capability discovery with offline registry support

Changed German synthesis to default to the verified Martin v1.2 fp32 profile with explicit legacy model compatibility

BREAKING: Changed to require explicit document languages and SSMD spans for mixed-language synthesis

Changed the locked dependency set to the v0.9 companion releases

Changed the pipeline to prepare frontend text once and render multiple lexicon variants through one acoustic backend

Changed audio results to expose prepared and source text while allowing nested language-span overrides

Changed German lexicon provisioning to support automatic or installed-only Lexphon data

Fixed German named-lexicon output to remain within the Kokoro phoneme vocabulary

Documented explicit language configuration, isolated example artifacts, and sequential example verification

Added registry-backed model discovery and verified distribution asset downloads

Added language profiles and split-layout Thai Wayu runtime support

Added maintained German model examples and a model and language registry inspection and synthesis example

Changed the pipeline to explicit Spokenform, Phrasplit, and prepared G2P stages with remapped offsets

Added ONNX input and waveform diagnostics with configurable invalid-output handling

Added named German G2P lexicon selection and profile-aware cache keys

Changed the minimum Spokenform dependency to 0.3.5 for the explicit text-preparation pipeline

Changed German examples for optional lexicon audio and currency normalization with cents

Fixed Python 3.10 compatibility when recording installed-model timestamps

Fixed parser metadata to preserve an existing spaCy sentence-model selection

Added model-derived word timings to audio results with clean-text offsets and sample positions

Raised the minimum PhraseSplit version to 0.3.6 and the SSMD version to 0.8.4

Plain-text sentence metadata now comes from the PhraseSplit operation that produced the document segments

Fixed timing alignment for named duration outputs and final-waveform transformations

Removed the separate plain-text PhraseSplit model-resolution preflight pass

Added low-startup-latency sentence streaming playback through one persistent bounded output stream

Added sentence audio units to the prepared-unit rendering API

Added model-derived WordTiming metadata to audio results for timestamp-capable Kokoro models

Fixed public timing offsets and fallback ownership to track the final returned waveform

Documented global document preparation with lazy sentence audio generation and streaming examples

Documented clean-text character offsets, sample-domain timing semantics, and unsupported waveform behavior

Added polynorm benchmark suite with evaluation harness, phoneme data generation, and CI workflow

Fixed polynorm benchmark evaluation and test harness stability

Added configurable emphasis gain scaling with numeric dB resolution and a bounded emphasis_gain_scale setting

Added a spoken-form showcase for raw structured text through the PyKokoro pipeline

Added configurable SSMD emphasis gain scaling for gain-only automatic emphasis approximation

Changed the kokorog2p minimum to 0.8.0 and enabled source-aware Spokenform preparation

Documented automatic kokorog2p spoken-form preparation separately from explicit SSMD say-as overrides

German synthesis defaults to the verified Martin v1.2 fp32 profile with automatic language-aware model and voice resolution

GitHub v1.0 model assets use the embedded vocabulary and no longer require a HuggingFace config download

Fixed managed HuggingFace voice loading to retain the canonical archive file path instead of the containing directory

Added paragraph-unit preparation and streaming for bounded, resumable document rendering

Added tri-state spaCy model policy with automatic local model discovery and with_spacy_model() helper

Changed documentation to pipeline-first API and archived legacy examples

Fixed SSMD pause precedence for explicit zero durations and voice-change boundaries

Added compact AudioResult retention controls and explicit audio-release methods

Added configurable AudioSig WSOLA, ESOLA, TD-PSOLA, and phase-vocoder prosody backends with a strict comparison harness

Changed pipeline result finalization to preserve markers while releasing optional segment audio

Changed Kokoro teardown to release dependent backend resources deterministically and idempotently

Changed combined speech prosody processing to one AudioSig compositor pass with WSOLA as the default and configured fallback behavior

Documented compact result memory behavior, ownership, release semantics, and the limits of non-streaming generation

Documented prosody backend constraints, experimental status, comparison workflow, and segment-level speech-quality limitations

Changed SSMD emphasis to metadata-preserving by default with opt-in audio approximation

Added explicit SSMD emphasis policies and diagnostics for plain, approximate, warning, and error handling

Fixed the tokenizer configuration example in the basic usage documentation

Improved user documentation examples and installation guidance across the supported workflows

Added default-safe SSMD 0.8 front-matter parsing with title and document metadata

Added portable provider voice bindings, deterministic pause-default precedence, marker offsets, resolver-based audio annotations, and explicit capability diagnostics

Raised the SSMD dependency window to >=0.8.0,<0.9 and documented literal-header migration

Added capability-driven ONNX execution-provider resolution for runtime-reported providers including NNAPI and XNNPACK

Added source-, variant-, and quality-aware model asset inspection APIs

Fixed Kokoro downloads and status checks to use exact model, config, and voice archive paths

Documented Android/Termux provider selection and dependency-light asset inspection

Added phrase-based short-sentence handling with energy-valley cutting for improved annunciation of short phrases

Added SHA256 verification, pinned HF revisions with offline mode, and stale download lock recovery

Changed SSMD syntax to [text]{key=”value”} annotation format, replacing legacy shorthand notations

Changed HuggingFace v1.0 ONNX repository to use timestamped model variant

Fixed Windows download cache path handling for model and voice files

Fixed voice name in English example script

Changed documentation format from reStructuredText to Markdown for improved readability and GitHub rendering

Added better support for spaCy models across all languages

Added automatic pause mode that adds pauses after sentences

Changed documentation with updated usage examples

Added German language support with ONNX model files

Added threading support for concurrent audio generation

Fixed Windows cache path handling and default backend selection

Fixed unit tests and mypy type errors

Added integrated splitter into doc parsers, removing separate splitter pipeline

Added automatic language detection based on voice selection

Added improved SSMD parsing with better span handling

Changed pipeline lifecycle and audio generator for better stability

Fixed paragraph splitting, pause parsing, and voice downloading

Added model_path and voices_path settings in pipeline

Fixed missing espeak and soundfile dependencies

Removed deprecated _phoneme_dictionary

Added new pipeline architecture with improved phoneme handling

Added improved PhonemeSegment with better prosody support

Added improved pipeline architecture with multiple fixes

Changed pipeline to drop legacy tokenizer in favor of new architecture

Changed all examples to use new pipeline format

Fixed phrasplit splitter and kokorog2p issues

Added audio transcript support for better unit testing and debugging

Changed SSMD parser to adapt to new phoneme syntax in kokorog2p

Fixed SSMD parser and detection logic

Added GenerationConfig dataclass to simplify TTS API

Added heading support from SSMD markup

Added configurable short sentence handler with multiple strategies

Fixed SSMD phoneme override bug in segment merging

Fixed short sentence handling and unit tests

Added SSMD syntax support with voice switching and prosody control

Added says-as implementation for cardinal, ordinal, and telephone numbers

Added optimal phoneme length calculation

Changed tokenizer and improved backend stability

Added model and voice download with manager classes for better code organization

Added configurable pause times for clause, sentence, and paragraph boundaries

Added session parameter configuration and pause variance support

Added HuggingFace Kokoro v1.1-zh model support

Added automatic splitting when max token limit is reached