Malaysian-TTS-0.6B-v1

Continue pretraining mesolitica/Malaysian-TTS-0.6B-v0.1 on much consistent dataset,

Use DistilCodec as speech detokenizer, output in 24k sample rate.
Support context switching between Malay and English.
Better pronunciation for letters.
Better repetitive tolerance.

Speakers

How do we train

Multipacking with proper document masking on 4096 context length.
FP32-BF16 mixed precision training.
Full parameter finetuning.
WanDB at https://wandb.ai/huseinzol05/Malaysian-TTS-0.6B-v1

How to use

First install DistilCodec,

pip3 install git+https://github.com/mesolitica/DistilCodec

Load the models,

# wget https://huggingface.co/IDEA-Emdoor/DistilCodec-v1.0/resolve/main/model_config.json
# wget https://huggingface.co/IDEA-Emdoor/DistilCodec-v1.0/resolve/main/g_00204000

from distilcodec import DistilCodec, demo_for_generate_audio_codes
from transformers import AutoTokenizer, AutoModelForCausalLM

codec_model_config_path='model_config.json'
codec_ckpt_path = 'g_00204000'

codec = DistilCodec.from_pretrained(
    config_path=codec_model_config_path,
    model_path=codec_ckpt_path,
    use_generator=True,
    is_debug=False).eval()

tokenizer = AutoTokenizer.from_pretrained('mesolitica/Malaysian-TTS-0.6B-v1')
model = AutoModelForCausalLM.from_pretrained('mesolitica/Malaysian-TTS-0.6B-v1', torch_dtype = 'auto').cuda()

Generate,

import soundfile as sf
from tqdm import tqdm

speakers = [
    'husein',
    'idayu',
    'singaporean',
    'DisfluencySpeech',
    'singlish-speaker2050',
    'singlish-speaker2202',
    'haqkiem',
]

string = 'IC saya adalah, sembilan enam, kosong tiga, satu empat, one, one, one, one, A, B, C, D, and yes, what can I help you sir?'

for s in tqdm(speakers):

    left = s +': ' + string
    prompt = f'<|im_start|>{left}<|speech_start|>'
    
    generate_kwargs = dict(
        **tokenizer(prompt, return_tensors = 'pt', add_special_tokens = False).to('cuda'),
        max_new_tokens=1024,
        temperature=0.7,
        do_sample=True,
        repetition_penalty=1.1,
    )
    generation_output = model.generate(**generate_kwargs)
    speech_token = tokenizer.decode(generation_output[0]).split('<|speech_start|>')[-1].replace('<|endoftext|>', '')
    numbers = re.findall(r'speech_(\d+)', speech_token)
    d = list(map(int, numbers))
    y_gen = codec.decode_from_codes(
        d,
        minus_token_offset=False
    )
    sf.write(f'{s}.mp3', y_gen[0, 0].cpu().numpy(), 24000)

Output,

Limitation

This model trained on normalized text, so if you have text such as 123, you have to normalize it first to become one two three or one hundred twenty three or satu dua tiga or seratus dua puluh tiga. Feel free to use Malaya for normalization, Malaya support Malay and English normalization, read more at https://github.com/mesolitica/malaya/issues/247#issuecomment-3030313021
The repetitive pronunciation dataset does not consistently use commas to for pauses. For example, A, A, A, A, B, B in our recordings is spoken as A A A A B B. We have no intention to improve it due to cost, but continue finetune using proper dataset should able to solve it.