initial

README.md

@@ -1,12 +1,8 @@
----
 title: Malay VITS
-emoji: 🐨
+emoji: 🗣️
 colorFrom: pink
-colorTo: gray
+colorTo: purple
 sdk: gradio
 sdk_version: 5.14.0
 app_file: app.py
-pinned: false
----
-
-Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference
+pinned: false

app.py

@@ -0,0 +1,91 @@
+import os
+
+os.environ['CUDA_VISIBLE_DEVICES'] = ''
+os.environ['TF_FORCE_GPU_ALLOW_GROWTH'] = 'true'
+
+from huggingface_hub import snapshot_download
+from malaya_speech.torch_model.vits.model_infer import SynthesizerTrn
+from malaya_speech.torch_model.vits.commons import intersperse
+from malaya_speech.utils.text import TTS_SYMBOLS
+from malaya_speech.tts import load_text_ids
+from malaya_speech.utils.astype import float_to_int
+import gradio as gr
+import torch
+import os
+import json
+
+try:
+    from malaya_boilerplate.hparams import HParams
+except BaseException:
+    from malaya_boilerplate.train.config import HParams
+
+speaker_id = {
+ 'Husein': 0,
+ 'Shafiqah Idayu': 1,
+}
+
+normalizer = load_text_ids(pad_to = None, understand_punct = True, is_lower = False)
+
+folder = snapshot_download(repo_id="malaysia-ai/malay-VITS-multispeaker")
+
+with open(os.path.join(folder, 'config.json')) as fopen:
+    hps = HParams(**json.load(fopen))
+
+model = SynthesizerTrn(
+    len(TTS_SYMBOLS),
+    hps.data.filter_length // 2 + 1,
+    hps.train.segment_size // hps.data.hop_length,
+    n_speakers=hps.data.n_speakers,
+    **hps.model,
+).eval()
+model.load_state_dict(torch.load(os.path.join(folder, 'model.pth'), map_location='cpu'))
+    
+
+def tts(text, speaker, temperature, length_ratio):
+    if len(text) < 1:
+        raise gr.Error('input text must longer than 0')
+
+    if speaker not in speaker_id:
+        raise gr.Error('speaker is not available')
+
+    t, ids = normalizer.normalize(text, add_fullstop = True)
+    if hps.data.add_blank:
+        ids = intersperse(ids, 0)
+    ids = torch.LongTensor(ids)
+    ids_lengths = torch.LongTensor([ids.size(0)])
+    ids = ids.unsqueeze(0)
+    sid = torch.tensor([speaker_id[speaker]])
+
+    with torch.no_grad():
+        audio = model.infer(
+            ids,
+            ids_lengths,
+            noise_scale=0.0,
+            noise_scale_w=0.0,
+            length_scale=1.0,
+            sid=sid,
+        )
+        y_ = audio[0].numpy()
+
+    data = float_to_int(y_[0, 0])
+    return (22050, data)
+
+demo = gr.Interface(
+    fn=tts,
+    inputs=[
+        gr.components.Textbox(label='Text'),
+        gr.components.Dropdown(label='Available speakers', choices=list(speaker_id.keys()), value = 'Husein'),
+        gr.Slider(0.0, 1.0, value=0.6666, label='temperature, changing this will manipulate pitch'),
+        gr.Slider(0.0, 3.0, value=1.0, label='length ratio, changing this will manipulate duration output'),
+    ],
+    outputs=['audio'],
+    examples=[
+        ['Syed Saddiq berkata, mereka seharusnya mengingati bahawa semasa menjadi Perdana Menteri Pakatan Harapan', 'Husein', 0.6666, 1.0],
+        ['Shah Alam - Pertubuhan Kebajikan Anak Bersatu Selangor bersetuju pihak kerajaan mewujudkan Suruhanjaya Siasatan Diraja untuk menyiasat isu kartel daging.', 'Shafiqah Idayu', 0.6666, 1.0],
+    ],
+    cache_examples=False,
+    title='End-to-End Malay TTS using VITS',
+)
+
+demo.queue().launch(server_name='0.0.0.0')
+    

requirements.txt

@@ -0,0 +1,7 @@
+torch==2.5.1
+torchaudio==2.5.1
+malaya-speech @ git+https://github.com/mesolitica/malaya-speech
+malaya
+malaya-boilerplate
+matplotlib
+transformers