piper-voices-thai

Piper-format Thai voices (converted from MMS-TTS)

Two Thai TTS voices packaged in Piper format (.onnx + .onnx.json), so they load directly with piper-tts — no separate inference engine needed.

Piper has no official Thai voice: rhasspy/piper-voices has no th/ directory. These are format conversions of existing MMS-TTS Thai fine-tunes.

Both source repos have 16 kHz siblings (FEMALEV1, MALEV2, MALE-NARRATOR); the two above were picked because they are 22.05 kHz, matching Piper's own voices.

Usage

import wave
from piper.voice import PiperVoice

voice = PiperVoice.load("th_TH-mms_female-medium.onnx")
with wave.open("out.wav", "wb") as f:
    voice.synthesize_wav("สวัสดีครับ นี่คือเสียงตัวอย่าง", f)

How the conversion works

MMS-TTS and Piper are both VITS, so only the ONNX graph signature and the text representation differ.

• The model is re-exported with Piper's input signature — input, input_lengths, scales — wrapping transformers' VitsModel.
• The config uses "phoneme_type": "text", where Piper treats each Unicode codepoint as a phoneme. That matches MMS exactly: a 71-character Thai vocabulary, no phonemization, no espeak-ng.
• Piper emits the id sequence BOS, PAD, c1, PAD, …, cN, PAD, EOS, while MMS was trained on intersperse(seq, 0) — 0, c1, 0, …, cN, 0. Mapping BOS and EOS to empty lists in phoneme_id_map makes the two line up exactly. PAD maps to [0], MMS's blank token.

Known limitation

The MMS vocabulary only contains the digits 0, 1, 2 and 4. Any other digit is dropped as an unknown character. Spell numbers out as words before synthesis.

Credits and license

cc-by-nc-4.0, inherited from facebook/mms-tts (the base model these fine-tunes derive from). Non-commercial use only.

• Base model — Meta AI, MMS-TTS (CC-BY-NC-4.0)
• Thai fine-tunes — VIZINTZOR
• Piper — OHF-Voice/piper1-gpl