uccharan (Python)

5× faster than espeak-ng. Just as accurate. Free to ship. A drop-in phonemiser for Piper, Kokoro and sherpa-onnx voices — English, Hindi, nineteen more South Asian languages, Spanish, French, German and Dutch, in Rust, Python, C and the browser. Apache-2.0. (measurements)

Permissively licensed text-to-phoneme conversion for neural TTS: English (US, British RP), Hindi and Hinglish, in the shapes Piper, Kokoro, Kitten TTS, Matcha-TTS and sherpa-onnx voices expect. No GPL code or data; no runtime dependencies.

pip install uccharan

import uccharan

uccharan.phonemize("Hello, world.")                          # 'həlˈoʊ, wˈɜːld.'
uccharan.phonemize("नमस्ते दुनिया", lang="hi")                  # 'nəmˈʌsteː dˈʊnɪjˌaː'
uccharan.phonemize("Hi there. Bye!", "en-us", "piper")       # ['hˈaɪ ðˈɛɹ.', 'bˈaɪ!']
uccharan.phonemize("Hi there.", "en-us", "sherpa-onnx")      # [[104, 712, 97, 618, …]]
uccharan.phonemize("I read it yesterday.", improved=True)    # 'aɪ ɹˈɛd ɪt jˈɛstɚdˌeɪ.'
uccharan.phonemize("I read the book.", "en-us", "kokoro")    # 'ˌI ɹˈɛd ðə bˈʊk.'  (misaki conventions, for Kokoro)
uccharan.phonemize_batch(lines, "hi")                        # all cores, GIL released

For code written against phonemizer:

from uccharan.compat import EspeakBackend
backend = EspeakBackend("en-us", preserve_punctuation=True, with_stress=True)
backend.phonemize(["Hello, world."], strip=True)             # ['həlˈoʊ, wˈɜːld.']

A command-line program is installed as uccharan (uccharan -l hi -p piper "नमस्ते").

For training voices and checking pronunciation, uccharan.train and the uccharan-train command label a dataset (training-ready phonemes for Piper, Matcha-TTS and icefall-style recipes, a symbol table and a voice.json), choose a reading script, report phoneme and diphone coverage, and score a recogniser's output against a reference word by word (guide).

English for Kokoro uses a port of misaki with misaki 0.9.4's dictionaries; load_misaki(dir) loads another copy.

The wheel (about 3 MB) carries the en-US, Hindi and other South Asian, Spanish, German, Dutch and French language data. British English and the Kokoro English data are language packs: downloaded from the GitHub release of the same version the first time they are needed, checked against SHA-256 sums inside the wheel, and cached in ~/.cache/uccharan. On an offline machine put the .ucd files (from the release page) in a directory named by UCCHARAN_DATA_DIR, set UCCHARAN_OFFLINE=1 to forbid downloads, or call uccharan.load_pack(path).

Documentation, measured agreement, speed and end-to-end results: https://github.com/cB-Sigma/uccharan.

Licence

Licence: Apache-2.0. Includes data from the CMU Pronouncing Dictionary (BSD-2-Clause) and misaki (Apache-2.0); see NOTICE.