NewYour coding agent can read the release notes before it upgrades.Set up the MCP server →
PyPI
PyPI, #2832 most downloaded on PyPI
Fast and local neural text-to-speech engine
Latest stable
1.8.0
Last release 2 months ago
04 Sep 2026
Ships unpredictably
gaps range from 3 weeks to 1.9 years
Most releases are documented
notes for 9 of 11 stable releases
Nothing withdrawn
no release was ever pulled
3 years old
11 releases · first in 2023
Drag across the ruler to compare, or tap a release to open it.
Add Thai phonemizer using TLTK in the new th extra
th extra
--data.phoneme_type thai for training; "phoneme_type": "thai" in a voice config for synthesisth_dict holds no lexicon, so unspaced Thai is never segmented; the leading vowels เ แ โ ใ ไ are not reordered; and a tone mark deletes the syllable's vowel, collapsing ป่า/ป้า/ป๊า/ป๋า to the same phonemesphonemize_chinese uses for Mandarin tonesscript/setup --th, and install the th extra in CI so the Thai tests runVitsModel sets automatic_optimization = False, and under manual optimization Lightning does not step the schedulers returned by configure_optimizers — nothing else did either, so the learning rate stayed pinned at its initial value for the whole runlr_schedulers last_epoch=0 no matter how many epochs it trainedon_train_epoch_end now steps both schedulers, and logs lr_g/lr_d--model.lr_decay/--model.lr_decay_d from --trainer.max_epochs when they are not set explicitly
0.999875, from upstream VITS) assumes a ~20,000-epoch schedule and decays the learning rate by only ~1% per 100 epochs, so simply stepping the scheduler would have changed almost nothing--model.lr_final_ratio (default 0.05) of the initial learning rate over the run; --model.lr_final_ratio 1.0 keeps it constantmax_epochs=-1, the default) have no run length to anneal over and still fall back to the upstream constants, with a warning--ckpt_path re-asserts the configured decay: ExponentialLR.state_dict() carries gamma, so the restored value would otherwise silently replace itLightningCLI._parse_ckpt_path merges a resumed checkpoint's saved hyper_parameters back over the parsed config, so --model.warmstart_ckpt comes back even when it is absent from the command lineon_fit_start and the optimizer/loop state after it, so the warmstart overwrote the resumed weights with the base model while keeping the restored epoch — discarding every epoch trained so faron_fit_start now skips both warmstart paths whenever trainer.ckpt_path is set, and logs that it didscript/libritts_r_to_csv, which converts extracted LibriTTS-R splits to Piper's metadata CSV
csv.writer(delimiter="|") so quoting round-trips through the csv.reader the trainer uses — a hand-rolled "|".join() silently drops the quotation marks that LibriTTS dialogue is full of.normalized.txt to the chapter .trans.tsv to .original.txt, and skips utterances whose audio has no transcript--min-seconds/--max-seconds filtering and a --stats duration histogramth extra
--data.phoneme_type thai for training; "phoneme_type": "thai" in a voice config for synthesisth_dict holds no lexicon, so unspaced Thai is never segmented; the leading vowels เ แ โ ใ ไ are not reordered; and a tone mark deletes the syllable's vowel, collapsing ป่า/ป้า/ป๊า/ป๋า to the same phonemesphonemize_chinese uses for Mandarin tonesscript/setup --th, and install the th extra in CI so the Thai tests runYour coding agent can read these notes before it upgrades. Set up the MCP server →
Add Japanese phonemizer using OpenJTalk (pyopenjtalk-plus) in the new ja extra
pyopenjtalk-plus) in the new ja extra
--data.phoneme_type japanese for training; "phoneme_type": "japanese" in a voice config for synthesisscript/setup --ja, and install the ja extra in CI so the Japanese tests runlibpiper: add piper_create_options and piper_create_with_options(), with piper_create() kept as a wrapper for ABI compatibilityRun the g2pW model through piper.g2pw_onnx instead of g2pw.api, dropping torch (~750 MB installed) and requests from the zh extra
piper.g2pw_onnx instead of g2pw.api, dropping torch (~750 MB installed) and requests from the zh extra
g2pw.api imports torch only to build padded tensors and iterate batches; the model itself already ran under onnxruntimeg2pW is still required, for its pinyin/bopomofo lookup tablesAdd Hebrew phonemizer using Nakdimon
Add libpiper C++ CLI executable ported from the legacy Piper repository, plus a C++ test suite
libpiper C++ CLI executable ported from the legacy Piper repository, plus a C++ test suitelibpiper builds on Windows (MSVC, MSYS2-GCC) and Windows CI--data.vowel_clusters)- Fix pathvalidate dependency
pathvalidate dependencyAdd missing wheels
Add Chinese phonemizer based on g2pW
quantize_dynamic--data.phoneme_type pinyin for Chinese phonemization using g2pW--data.phoneme_type text for using IPA phonemes directly (no espeak-ng)--model.vocoder_warmstart_ckpt <CHECKPOINT> to restore vocoder params only--data.dataset_type 'phoneme_ids' to train with pre-generated phoneme ids
--data.num_symbols <N> to set number of phonemes--data.phonemes_path "/path/to/phonemes.json" for phoneme/id map--output-dir-naming option with timestamp (default) and textMoved development to OHF-Voice org
libpiper written in C++ is plannedpiper-phonemize libraryPiperVoice.synthesize takes a SynthesisConfig and generates AudioChunk objectsPiperVoice.synthesize_raw is removedpiper.download_voices utility for downloading voices from HuggingFacepiper ... -- "Text to speak"--input-file <FILE>ffplay[[ <phonemes> ]]--volume <MULTIPLIER> (default is 1.0)Nothing published for this version
Nothing published for this version
One column per quarter.