Update app.py

app.py

@@ -90,11 +90,71 @@ def call_generate_text(message, history):
     return "",history
 
 
-with gr.Blocks(title="LLM with TTS") as demo:
+
+head = '''
+<script src="https://cdn.jsdelivr.net/npm/onnxruntime-web/dist/ort.webgpu.min.js" ></script>
+    
+    <script type="module">
+        
+        import { MatchaTTSRaw } from "https://akjava.github.io/Matcha-TTS-Japanese/js-esm/matcha_tts_raw.js";
+        import { webWavPlay } from "https://akjava.github.io/Matcha-TTS-Japanese/js-esm/web_wav_play.js";
+        import { arpa_to_ipa } from "https://akjava.github.io/Matcha-TTS-Japanese/js-esm/arpa_to_ipa.js";
+        import { loadCmudict } from "https://akjava.github.io/Matcha-TTS-Japanese/js-esm/cmudict_loader.js";
+        import { env,textToArpa} from "https://akjava.github.io/Matcha-TTS-Japanese/js-esm/text_to_arpa.js";
+        
+        env.allowLocalModels = true;
+        env.localModelPath = "./models/";
+        env.backends.onnx.logLevel = "fatal";
+        
+        let matcha_tts_raw;
+        let cmudict ={};
+        let speaking = false;
+        async function main(text,speed=1.0,tempature=0.5,spk=0) {
+            console.log(text)
+            if (speaking){
+                console.log("speaking return")
+            }
+            speaking = true
+            console.log("main called")
+            if(!matcha_tts_raw){
+                matcha_tts_raw = new MatchaTTSRaw()
+                console.time("load model");
+                await matcha_tts_raw.load_model('https://huggingface.co/spaces/Akjava/matcha-tts-onnx-benchmarks/resolve/main/models/matcha-tts/ljspeech_sim.onnx',{ executionProviders: ['webgpu','wasm'] });
+                
+                console.timeEnd("load model");
+                
+                let cmudictReady = loadCmudict(cmudict,'https://akjava.github.io/Matcha-TTS-Japanese/dictionaries/cmudict-0.7b')
+                await cmudictReady
+            }else{
+                console.log("session exist skip load model")
+            }
+           
+            const arpa_text = await textToArpa(cmudict,text)
+            const ipa_text = arpa_to_ipa(arpa_text).replace(/\s/g, "");
+            console.log(ipa_text)
+            const spks = 0
+            
+            console.time("infer");
+            const result = await matcha_tts_raw.infer(ipa_text, tempature, speed,spks);
+            
+            if (result!=null){
+                console.timeEnd("infer");
+                webWavPlay(result)
+               
+            }
+    
+            speaking = false
+        }
+        window.MatchaTTSEn = main
+        console.log(MatchaTTSRaw)
+</script>
+'''
+with gr.Blocks(title="LLM with TTS",head=head) as demo:
     gr.Markdown("LLM and TTS models will change without notice.")
     js = """
     function(chatbot){
-    console.log(chatbot[chatbot.length -1])
+    text = (chatbot[chatbot.length -1])["content"]
+    window.MatchaTTSEn(text)
     }
     """
     chatbot = gr.Chatbot(type="messages")