Revolab VITS — Multi-Speaker Bahasa Melayu TTS

Piper TTS voice models trained on Revolab Malay speech datasets using VITS architecture.

Speakers (56 trained)

Production Quality (CER < 10%)

Speaker ID	Name	Samples	CER	WER
sarah	sarah	27,792	0.0402	0.1856
pendakwah	pendakwah	46,222	0.0619	0.1847
pendakwah_teknologi	Pendakwah Teknologi	46,222	0.0619	0.1847
paan	Paan	27,434	0.0630	0.1759
anwar	AI	100	0.0732	0.2101

Usable Quality (CER 50-65%)

Speaker ID	Name	Samples	CER	WER
G0095_0001	G0095	12,317	0.5382	0.7685
1614-NorinaYahya	Noriya	27,962	0.5490	0.8351
8-enhanced-v2	Sarah	27,991	0.6084	0.9322
Iqbal_25	Iqbal	11,077	0.6107	0.8428
JackLim_1	Jacky	9,143	0.6390	0.9050

Low Quality (CER > 75%)

Speaker ID	Name	Samples	CER	WER
Angry_2	Pearl (Angry 2)	27,999	0.7556	0.9965
cv_971b3c0c6dbd	cv_971b3c0c6dbd	6,632	0.7557	0.9845
Surprise	Surprise	5,986	0.7604	0.9998
G0068_0003	G0068	12,657	0.7670	0.9976
G0004_0012	G0004	12,656	0.7756	1.0228
G0016_0001	G0016	12,694	0.7760	1.0000
cv_06a7ed020ffa	cv_06a7ed020ffa	7,050	0.7771	0.9975
Dania_22	Dania	13,558	0.7777	0.9869
2394-Digi-Mohon-Share-Amel-BM	Amel	29,273	0.7784	1.0086
cv_0c744927a83d	cv_0c744927a83d	7,461	0.7786	0.9988
cv_e6ca744ffbb2	cv_e6ca744ffbb2	6,604	0.7813	0.9869
cv_1a952f7ad4bd	cv_1a952f7ad4bd	6,587	0.7829	1.0155
berani_buat	berani_buat	8,657	0.7853	0.9967
Happy	Pearl (Happy)	12,317	0.7854	1.0000
Sad	Pearl (Sad)	7,202	0.7855	0.9907
k-lao-ZH	k lao ZH	5,533	0.7875	1.0042
cv_a2b722b2d474	cv_a2b722b2d474	5,805	0.7887	0.9907
Sabil_2	Kamal	8,276	0.7918	1.0093
cv_51062240a78a	cv_51062240a78a	6,929	0.7944	1.0170
mych_male_yt_13887	mych male yt 13887	5,967	0.7954	1.0098
sugumaran_6	sugumaran 6	5,630	0.7969	1.0052
Alvin_6	Alvin	27,996	0.7989	1.0199
cv_9d9664719e9d	cv_9d9664719e9d	7,020	0.8016	1.0319
Angry_1	Pearl (Angry 1)	27,849	0.8026	0.9905
cv_bd96fb646e95	cv_bd96fb646e95	6,552	0.8061	0.9896
cv_c0b0ce94caa2	cv_c0b0ce94caa2	7,053	0.8104	1.0093
niketh_23	niketh 23	5,767	0.8177	1.0074
angellyn	angellyn	6,096	0.8185	1.0000
hendrick-22	Hendrick	6,774	0.8185	0.9880
Farid_596	Faiz	5,216	0.8574	0.9966

Non-Malay Speakers (no eval)

Speaker ID	Name	Samples	Language
Akari_Kitou	Akari Kitou	6,772	Japanese
Aoi_Yuuki	Aoi Yuuki	6,772	Japanese
Ayana_Taketatsu	Ayana Taketatsu	10,970	Japanese
Celine-ZH	Celine Zh	10,967	Chinese
ChangYong	Changyong	16,195	Chinese
Eri_Kitamura	Eri Kitamura	5,209	Japanese
Haruka_Tomatsu	Haruka Tomatsu	6,758	Japanese
Kana_Hanazawa	Kana Hanazawa	8,778	Japanese
Maaya_Uchida	Maaya Uchida	8,368	Japanese
Nao_Touyama	Nao Touyama	8,400	Japanese
Rie_Kugimiya	Rie Kugimiya	15,001	Japanese
Rie_Takahashi	Rie Takahashi	13,534	Japanese
Rina_Satou	Rina Satou	12,969	Japanese
Sora_Amamiya	Sora Amamiya	6,901	Japanese

Evaluation Methodology

Evaluated using WhisperX large-v3 transcription + revo-norm text normalization on 36 sample texts covering:

• Digits/currency (RM10.50, 03-8888 9999)
• Malaysian locations (Jalan Ampang, Pulau Pinang)
• Names — Malay/Chinese/Indian (Encik Ahmad, Lim Wei Jie, Rajesh Kumar)
• English code-switching (Meeting kita pukul 3 petang)
• Dates/time (30hb Jun 2024, pukul 8 pagi)
• Addresses (25, Jalan Taman Indah 3/4, 43000 Kajang)
• Financial terms (Baki akaun RM5,670.23)
• General conversation

CER = Character Error Rate (lower is better). WER = Word Error Rate.

Full per-speaker results: eval_all/<speaker>_eval.csv

Install

pip install piper-tts
pip install git+https://github.com/khursanirevo/revo-norm.git

Quick Start

from piper import PiperVoice
from revo_norm import normalize_text

raw = "Harga RM10.50 sahaja"
text = normalize_text(raw, language="ms")

voice = PiperVoice.load("hf_repo/speakers/sarah/model.onnx",
                         config_path="hf_repo/speakers/sarah/model.onnx.json")
audio = voice.synthesize(text)

Text Normalization

revo-norm handles Malay-specific normalization:

• Numbers → words ("115" → "seratus lima belas")
• Currency ("RM10.50" → "sepuluh ringgit lima puluh sen")
• Dates, phone numbers, percentages, temperatures, etc.

Structure

speakers.json              # Speaker registry with eval metrics
speakers/
  <name>/
    model.onnx             # ONNX export for inference
    model.onnx.json        # Phoneme config

Performance (CPU)

Metric	Value
Avg latency	~54ms
Avg RTF	0.030
Speed	33.6x realtime

Training

All models trained with:

• Architecture: VITS
• Phonemizer: espeak-ng (ms voice)
• Sample rate: 22050Hz
• GPU: NVIDIA H200 NVL