v0.27.0

What's Changed

Features

• Speaker caching for cloned voices by @eginhard in #438
  • For usage details see https://coqui-tts.readthedocs.io/en/latest/cloning.html
  • ⚠️ The old caching mechanism of Bark and Tortoise has been removed, switch to the new one instead.
• Provide synthesize() method with a common interface for every TTS model by @eginhard in #453
  • ⚠️ Deprecate speaker_id argument of synthesize(), use speaker instead.
  • ⚠️ Deprecate config argument of synthesize(), it can safely be left out.
• Add OpenAI-compatible endpoint to the server by @teddybear082 in #421
  See https://coqui-tts.readthedocs.io/en/latest/server.html#openai-compatible-endpoint

Fixes

• Update coqui-tts-trainer to 0.3.0 to fix numerous training-related bugs by @eginhard in #423
  For the full list of fixes see https://github.com/idiap/coqui-ai-Trainer/releases/tag/v0.3.0
• fix(configs): add default padding character by @eginhard in #425
• Fix KeyError when speaker_id is empty string in TTS server by @mehulanshumali in #436
• Fix: Update xtts finetuning Colab to support Gradio 5 by @eulphean in #408
• Bring back compatibility with numpy1 by @MarwanMashra in #413
• fix: update XTTS/Tortoise GPT code for HF transformers 4.52+ by @eginhard in #414
• build: lower minimum pytorch version back to 2.1 by @eginhard in #432
• refactor(phonemizer): replace mecab-python3 with fugashi for japanese by @eginhard in #417
• docs: add Docker Compose config by @KishoOoOoOo in #411

New Contributors

• @KishoOoOoOo made their first contribution in #411
• @eulphean made their first contribution in #408
• @MarwanMashra made their first contribution in #413
• @teddybear082 made their first contribution in #421
• @mehulanshumali made their first contribution in #436

Full Changelog: v0.26.2...v0.27.0

v0.26.2

What's Changed

Fixes

• fix(xtts): restrict transformers to <4.52 to avoid corrupted output by @eginhard in #396
• fix(xtts): fix Colab demo package installation by @eginhard in #396
• fix(xtts): provide more helpful error message when reference audio is too short by @eginhard in #396
• fix: don't convert to int to avoid constant value in onnx exports by @eginhard in #390

Full Changelog: v0.26.1...v0.26.2

v0.26.1

What's Changed

Features

• feat(server): support multi-speaker models in MaryTTS endpoint by @Sleuth56 in #351

Fixes

• Switch to Numpy>=2, Pytorch>=2.3 by @fabiocat93 in #346
• fix(xtts): update colab finetuning notebook by @eginhard in #377
• fix(forward_tts): ensure tensor 'g' is on the same device as 'x' by @btseee in #378
• Remove Spacy dependency by @eginhard in #383

New Contributors

• @fabiocat93 made their first contribution in #346
• @btseee made their first contribution in #378
• @Sleuth56 made their first contribution in #351

Full Changelog: v0.26.0...v0.26.1

v0.26.0

What's Changed

Features

• Added speaker_wav parameter to the server by @shavit in #295
• feat(api): support setting speed by @eginhard in #316
• Added new persian-tts-female-vits model by @DrewThomasson in #332

Fixes

• docs: clean up server README by @junland in #272
• fix: notify users when wrong coqpit package is installed by @eginhard in #294
• Refactor for compatibility with transformers>=4.47 by @JohnnyStreet and @eginhard in #319
• Support use of --continue_path to resume XTTS training by @eginhard in #270
• Drop Python 3.9 support by @eginhard in #255

Dev

• Switch remaining CLI tests to Python, separate integration tests by @eginhard in #276
• Added paths-ignore in workflows by @DrewThomasson in #334

New Contributors

• @junland made their first contribution in #272
• @DrewThomasson made their first contribution in #332

Full Changelog: v0.25.3...v0.26.0

v0.25.3

What's Changed

Fixes

• fix(fairseq): handle change of model file name by @eginhard in #264

Full Changelog: v0.25.2...v0.25.3

v0.25.2

What's Changed

⚠️ Fairseq Vits models are broken in this release.

Features

• Add kNN-VC model by @eginhard in #256
• Support all Coqui TTS models in the server by @eginhard in #252
• Allow both Path and strings where possible and add type hints by @eginhard in #210
• feat(manager): print download location when listing models by @eginhard in #213

Fixes

• fix(bark): handle broken paths in config by @eginhard in #253
• fix(openvoice): correctly set utterance length by @eginhard in #260
• fix(bin): log to stdout in cli tools by @eginhard in #217
• fix(vc): support both cpu and cuda by @eginhard in #244
• fix(xtts): voice_dir should remain None if not specified by @eginhard in #224
• Fix num2words call using non-standard lang code by @SkaceKamen in #237
• chore: remove unused callback code by @eginhard in #229
• fix: convert >35 digit English numbers digit-by-digit by @lostways in #240
• Change old docker image url to the one that is relevant to this repo in README.md by @DelovoiDC in #243
• test: switch from nose2 to pytest by @eginhard in #208
• Update plot_embeddings_umap notebook by @eginhard in #221
• Improve documentation by @eginhard in #207

New Contributors

• @SkaceKamen made their first contribution in #237
• @lostways made their first contribution in #240
• @DelovoiDC made their first contribution in #243

Full Changelog: v0.25.1...v0.25.2

WavLM-HiFiGAN vocoders from kNN-VC

• HiFiGAN vocoders for WavLM features trained on LibriSpeech100 from https://github.com/bshall/knn-vc (MIT license)

v0.25.1

What's Changed

Features

• Expand Python API capabilities by @eginhard in #197

Fixes

• Fix XTTS voice cloning by @eginhard in #199
• fix(xtts): clearer error message when file given to checkpoint_dir by @eginhard in #184

Full Changelog: v0.25.0...v0.25.1

v0.25.0

What's Changed

⚠️ XTTS voice cloning is broken in this release.

Features

• Add OpenVoice VC models by @eginhard and @ajk1402 in #183

Fixes

• Automatically convert audio to mono, add more helpful error messages by @eginhard in #166
• fix(bin.synthesize): return speakers names only by @shavit in #147
• Show original model URLs by @eginhard in #149
• Support for building Docker on arm64 by @hongkongkiwi in #159
• refactor: handle deprecation of torch.cuda.amp.autocast by @eginhard in #144

Dev

• build: move doc dependencies from extra into group and build with uv by @eginhard in #133
• Use external package for monotonic alignment search by @eginhard in #135
• ci: allow testing out trainer/coqpit branches before release by @eginhard in #168
• Remove unused code by @eginhard in #172
• build: switch to forked coqpit by @eginhard in #110

New Contributors

• @hongkongkiwi made their first contribution in #159
• @ajk1402 made their first contribution in #183

Full Changelog: v0.24.3...v0.25.0

v0.24.3

What's Changed

Fixes

• Load weights only in torch.load for pytorch>=2.4 by @shavit in #77 and @eginhard in #113
• Add compatibility with transformers>=4.43 by @JohnnyStreet in #109
• fix(gpt): set attention mask and address other warnings by @eginhard in #114
• fix(text.characters): add nasal diacritic by @eginhard in #127

New Contributors

• @JohnnyStreet made their first contribution in #109

Full Changelog: v0.24.2...v0.24.3