Welsh/English Bilingual Neural TTS (Bangor)

A bilingual Welsh (Cymraeg) + English neural text-to-speech voice (VITS / Piper architecture) with three professional studio voices. It pronounces English natively (not Welsh-accented) and handles code-switched cy/en text within a sentence, because it is phonemized with the Bangor Welsh G2P rather than eSpeak.

Developed by: Language Technologies Unit (Uned Technolegau Iaith), Bangor University.

Two models, two licences — choose what you need

This repo contains two models sharing the same three voices; only the warm-start base differs:

Both are fine-tuned on the same Welsh/English studio recordings; only the warm-start base differs. The default gives slightly better English; the -cc0- file is for users who must avoid any attribution requirement.

Model details

• Architecture: Piper VITS, medium quality
• Speakers: 3 (sid 0, 1, 2)
• Sample rate: 22,050 Hz · Symbols: 256
• Phonemizer: Bangor Welsh G2P — backend: bangor, english_mode: native, data_version: 718542836d1dcbdd
• Punctuation-aware pauses; no eSpeak/GPL at inference.

⚠️ Important: phonemization

This model was trained with the Bangor G2P, so it will not pronounce correctly through the default eSpeak-based piper command. You must phonemize with the matching front-end:

• Python: pip install techiaith-g2p>=1.4.0 — BangorG2P(english_mode="native"). Source: https://github.com/techiaith/techiaith-g2p · CC0-1.0 · no dependencies
• C runtime: libcy_phonemize, in the same repository

Both emit the exact phoneme-id sequence the model expects; the runtime asserts data_version == 718542836d1dcbdd.

Usage (Python + ONNX Runtime + Bangor G2P)

pip install techiaith-g2p onnxruntime numpy

import numpy as np, onnxruntime as ort, wave
from techiaith.g2p import BangorG2P

g2p = BangorG2P(english_mode="native")
sess = ort.InferenceSession("cy_en_GB-bu_tts-bangor-medium.onnx")   # or ...-cc0-medium.onnx

def synth(text, sid=0, path="out.wav", noise_w=0.8):
    # For isolated letters / digits / very short "keyboard-echo" inputs, use noise_w≈0.3.
    ids = g2p.text_to_ids(text)
    scales = np.array([0.667, 1.0, noise_w], dtype=np.float32)  # noise, length, noise_w
    audio = sess.run(None, {
        "input": np.array([ids], dtype=np.int64),
        "input_lengths": np.array([len(ids)], dtype=np.int64),
        "scales": scales,
        "sid": np.array([sid], dtype=np.int64),
    })[0].squeeze()
    audio = (audio * (32767 / max(0.01, np.max(np.abs(audio))))).astype("<i2")  # peak-normalize
    with wave.open(path, "wb") as w:
        w.setnchannels(1); w.setsampwidth(2); w.setframerate(22050); w.writeframes(audio.tobytes())

synth("Bore da. Sut wyt ti'n teimlo heddiw?", sid=0)          # Welsh
synth("Good morning. How are you feeling today?", sid=1)      # English
synth("Nes i fwcio'r tickets ar-lein, so we're all set.", 2)  # code-switched

Model files

• cy_en_GB-bu_tts-bangor-medium.onnx (+ .onnx.json) — default model (CC-BY)
• cy_en_GB-bu_tts-bangor-cc0-medium.onnx (+ .onnx.json) — CC0 variant

Intended use & limitations

Screen readers and assistive technology, Welsh/bilingual language-learning apps, and local/offline TTS. Best on general conversational Welsh, English, and mixed cy/en text. Medium quality (assistive-grade, not studio mastering); OOV/technical terms, rare place names and unusual spellings may mispronounce; neutral tone, limited emotional range.

Licence & attribution

This repo is licensed CC-BY-4.0, reflecting the default model (cy_en_GB-bu_tts-bangor-medium). The cy_en_GB-bu_tts-bangor-cc0-medium file is available under CC0-1.0 (public domain, no attribution).

The studio voices, integration code, scripts and Bangor G2P are in-house contributions of the Language Technologies Unit (CC0). The default model was warm-started from Piper's en_GB-vctk-medium voice, so it is a derivative work of:

CSTR VCTK Corpus (v0.92). Yamagishi, Junichi; Veaux, Christophe; MacDonald, Kirsten (2019). The Centre for Speech Technology Research (CSTR), University of Edinburgh. https://doi.org/10.7488/ds/2645 — under the Open Data Commons Attribution License (ODC-By) v1.0 / CC BY 4.0. Piper voices © Michael Hansen (rhasspy), CC BY 4.0.

The -cc0- variant was not warm-started from VCTK: it uses en_US-ljspeech-medium (public domain), so it carries no attribution requirement.