← Yakthung (Limbu) tools

Limbu voice work in progress

A neural text-to-speech voice for Limbu (Yakthung) — the first we know of. Trained from timestamped, cleaned Limbu audio with a Limbu-specific Devanagari→phoneme frontend (glottal stop, vowel length, the lower-mid vowels). These are real renders from the model below; quality is early and improving.

Type Limbu, hear it live

Type Limbu in Devanagari and the model reads it aloud. First use may take ~30–60 s while the voice wakes up.

Or listen to fixed examples

Each line shows the Limbu text, the phonemes our frontend produced (note ː = long vowel, ʔ = glottal stop), and the synthesized audio.

इङ्‌गाॽ कुइसिःक् खे़ङ्‌हाॽआङ् इक्‌सादिङ् खाम्‌बेःक्‌मोबा मे़ःन्‌छिरो॥

i ŋ g a ʔ · k u i s i ː k · kh ɛ ŋ h a ʔ a ŋ · i k s a d i ŋ · kh a m b e ː k m o b a · m ɛ ː n ch i r o
22.05 kHz · best checkpoint (epoch 19) · speaker limbu_bible_voice

आनिगे़ सःप्‌मान् कुलिङ्‌धो के़त्‌ल फाॽआङ् कन् पाःन्‌हाॽ साप्‌तुम्‌बे़आङ् हाक्‌कासिगे़बारो॥

a n i g ɛ · s ɔ ː p m a n · k u l i ŋ dh o · k ɛ t l ɔ · ph a ʔ a ŋ · k ɔ n · p a ː n h a ʔ · s a p t u m b ɛ a ŋ · h a k k a s i g ɛ b a r o
same voice, a longer line · 22.05 kHz
a second render of this line (different inference settings)
Tip: the fixed examples above always play instantly (no model wake-up), so they're handy as a reliable fallback if the live widget is still warming up.

How it was built

Model Piper+ VITS fine-tune
Audio ~18.4 h, 22.05 kHz
Frontend Limbu Devanagari → phonemes
Best ckpt epoch 19, step 19,560
Speakers one Limbu voice
Status early / improving

Model card & checkpoints: ampixa/limbu-piper-lifwbt on Hugging Face. The voice is a research prototype; phonetic detail and naturalness are still being tuned with native review.