Kecil-TTS Michael v0.2

A Malay text-to-speech model with a male voice, 8.5 MB. One ONNX file, one voice (close to Kokoro's am_michael), runs in a browser tab (transformers.js, WebGPU or WASM) or on one CPU thread in Python.

This is the male sibling of Kecil-TTS Michael v0.3 (female voice). Same size, same code, same Malay front end.

First release; read the numbers below. It is understandable, but it makes more wrong words than the female model.

Model file onnx/model_quantized.onnx, 8.46 MB (int8 weights)
Parameters 7.9 M
Output 24 kHz mono
Language Bahasa Melayu (Malaysia), with common English words used in Malaysian speech
Inputs input_ids (int64 [1,T]) and speed (float32 [1])

Quick start

pip install onnxruntime numpy huggingface_hub soundfile
pip install https://huggingface.co/Suarify/kecil-tts-michael-v0.2/resolve/main/dist/kecil_tts_8m-0.3.0-py3-none-any.whl
import soundfile as sf
from kecil_tts import SAMPLE_RATE, KecilTTS

tts = KecilTTS.from_pretrained("Suarify/kecil-tts-michael-v0.2")
sf.write("hello.wav", tts("Selamat pagi, boleh saya bantu encik hari ini?"), SAMPLE_RATE)

examples/ has Python, Node and streaming scripts (stream_python.py, stream_node.mjs); js/ is the transformers.js front end.

How good is it?

Quality was measured by playing the model's speech to Whisper large-v3 and comparing with the text (character error rate, CER; lower is better; no model scores 0 because the recogniser itself misreads some speech).

Sentences the model had not seen n Median CER Capped mean CER Share at or under 8%
Test sentences, every 5th row of the 200-sentence test set (all six categories) 40 6.2% 10.2% 57%
Held-out sentences from the training set 60 6.8% 12.9% 52%

Test sentences by category:

Category n Median CER
codeswitch 4 13.1%
conversational 12 7.0%
general 14 8.2%
numbers 6 0.0%
question 2 0.0%
short 2 13.5%

For reference, the speech this model learned from scores 4.3% median CER on the held-out clips.

What changed since v0.1

The decoder now uses 128 sine harmonics instead of 64. For a 120 Hz voice 64 harmonics stop at about 7.5 kHz, so the rest of the spectrum was filled with noise (hiss); with 128 the 7.5-11 kHz band matches the training audio. Error rates are about the same; the sound is slightly cleaner but still not natural.

Limitations (read before using)

  • Error rate: 43% of the test sentences scored above 8% CER (a wrong or missing word or more); names, numbers, English words and long sentences are the weak spots. The female v0.3 model is more accurate.
  • Sound quality: it still sounds somewhat robotic or damp in places (heard by the maintainer on held-out sentences). The training audio is itself generated and voice-converted, so the student cannot copy it as crisply as a model distilled from a deterministic teacher.
  • One synthetic male voice, no emotion or style control. Some sentences sound slightly American-accented.
  • Trained on a small set: about 49 minutes of synthetic Malay speech (Chatterbox Multilingual, voice-converted twice: once to a female voice, then to the male voice).
  • Written for short sentences. Long text is split at sentence ends; the Python and JS front ends also support streaming.
  • It is synthetic speech. Please say so when you use it.

How it was made

  1. A rule-based Malay front end (kecil_tts/, ported line for line to js/) writes Malay as phonemes in the 178-symbol vocabulary of Kokoro-82M.
  2. About 750 Malay sentences were turned into speech with Chatterbox Multilingual and voice-converted to the male voice using a reference read by Kokoro am_michael. Audio after the last aligned word is cut.
  3. The 8M-parameter architecture and training recipe are those of Paradee-8M, starting from its male US voice row; under one hour on one T4 GPU.
  4. Weights were exported to ONNX (opset 17) with int8 weights and made WebGPU-safe.

License and credits

Code and weights: Apache-2.0. kecil_tts/lexicon_e.json and js/g2p/lexicon_e.json are derived from Wiktionary data and are CC BY-SA 4.0. See NOTICE: Paradee-8M (Sahil Mahendrakar), Kokoro-82M (hexgrad), Chatterbox Multilingual (Resemble AI), misaki, WikiPron, g2p-id, Tatoeba and Malay Wikipedia.

Downloads last month
-
Inference Providers NEW
This model isn't deployed by any Inference Provider. πŸ™‹ Ask for provider support

Model tree for Suarify/kecil-tts-michael-v0.2

Quantized
(11)
this model

Space using Suarify/kecil-tts-michael-v0.2 1