update

app.py

@@ -12,15 +12,12 @@ from fairseq.tasks.hubert_pretraining import LabelEncoder
 from fairseq.data.audio.speech_to_text_dataset import get_features_or_waveform 
 
 device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
-
-WORK_DIR = os.getcwd()
 checkpoint = torch.load('ckpts/clartts_tts.pt')
 checkpoint['cfg']['task'].t5_task = 't2s'
-checkpoint['cfg']['task'].hubert_label_dir = "utils/"
 checkpoint['cfg']['task'].bpe_tokenizer = "utils/arabic.model"
 checkpoint['cfg']['task'].data = "utils/"
-checkpoint['cfg']['model'].mask_prob = 0.0
-checkpoint['cfg']['task'].mask_prob = 0.0
+checkpoint['cfg']['model'].mask_prob = 0.5
+checkpoint['cfg']['task'].mask_prob = 0.5
 task = ArTSTTask.setup_task(checkpoint['cfg']['task'])
 
 emb_path='embs/clartts.npy'
@@ -56,20 +53,40 @@ def inference(text, spkr=emb_path):
         )
     with torch.no_grad():
         gen_audio = vocoder(outs.to(device))
-    return (16000,gen_audio.cpu().numpy())
+    speech = (gen_audio.cpu().numpy() * 32767).astype(np.int16)
+    return (16000,speech)
 
 text_box = gr.Textbox(max_lines=2, label="Arabic Text", rtl=True)
 out = gr.Audio(label="Synthesized Audio", type="numpy")
 title="ArTST: Arabic Speech Synthesis"
 description="ArTST: Arabic text and speech transformer based on the T5 transformer. This space demonstarates the TTS checkpoint finetuned on \
-    the CLARTTS dataset. The model is pre-trained on the MGB-2 dataset.Check the  <a href='https://github.com/mbzuai-nlp/ArTST'> ArTST repo</a> for implementation code and \
-    Read our <a href='https://arxiv.org/abs/2310.16621'>paper</a> for more details."
+    the Classical Arabic Text-To-Speech (CLARTTS) dataset. The model is pre-trained on the MGB-2 dataset."
+
 examples=["لأن فراق المألوف في العادة ومجانبة ما صار متفقا عليه بالمواضعة",\
     "ومن لطيف حكمته أن جعل لكل عبادة حالتين",\
     "فمن لهم عدل الإنسان مع من فوقه"]
 
+article = """
+<div style='margin:20px auto;'>
+<p>References: <a href="https://arxiv.org/abs/2310.16621">ArTST paper</a> |
+<a href="https://github.com/mbzuai-nlp/ArTST">GitHub</a> |
+<a href="https://huggingface.co/MBZUAI/ArTST">Weights and Tokenizer</a></p>
+<pre>
+@misc{toyin2023artst,
+      title={ArTST: Arabic Text and Speech Transformer}, 
+      author={Hawau Olamide Toyin and Amirbek Djanibekov and Ajinkya Kulkarni and Hanan Aldarmaki},
+      year={2023},
+      eprint={2310.16621},
+      archivePrefix={arXiv},
+      primaryClass={cs.CL}
+}
+</pre>
+<p>Speaker embeddings were generated from <a href="http://www.festvox.org/cmu_arctic/">CMU ARCTIC</a>.</p>
+</div>
+"""
+
 demo = gr.Interface(inference, \
-    inputs=text_box, outputs=out, title=title, description=description, examples=examples)
+    inputs=text_box, outputs=out, title=title, description=description, examples=examples, article=article)
 
 if __name__ == "__main__":
     demo.launch(share=True)