app.py

import gradio as gr
import requests
import uuid
import os
ASR_API = "http://astarwiz.com:9998/asr"
TTS_SPEAK_SERVICE = 'http://astarwiz.com:9603/speak'
TTS_WAVE_SERVICE = 'http://astarwiz.com:9603/wave'

LANGUAGE_MAP = {
    "en": "English",
    "ma": "Malay",
    "ta": "Tamil",
    "zh": "Chinese"
}

# Add a password for developer mode
DEVELOPER_PASSWORD = os.getenv("DEV_PWD")

def inference_via_llm_api(input_text, min_new_tokens=2, max_new_tokens=64):
    print(input_text)
    one_vllm_input = f"<|im_start|>system\nYou are a translation expert.<|im_end|>\n<|im_start|>user\n{input_text}<|im_end|>\n<|im_start|>assistant"
    vllm_api = 'http://astarwiz.com:2333/' + "v1/completions"
    data = {
        "prompt": one_vllm_input,
        'model': "./Edu-4B-NewTok-V2-20240904/",
        'min_tokens': min_new_tokens,
        'max_tokens': max_new_tokens,
        'temperature': 0.1,
        'top_p': 0.75,
        'repetition_penalty': 1.1,
        "stop_token_ids": [151645, ],
    }
    response = requests.post(vllm_api, headers={"Content-Type": "application/json"}, json=data).json()
    print(response)
    if "choices" in response.keys():
        return response["choices"][0]['text'].strip()
    else:
        return "The system got some error during vLLM generation. Please try it again."
    
def transcribe_and_speak(audio, source_lang, target_lang):
    if not audio:
        return "Please provide an audio input.", None, None

    # ASR
    file_id = str(uuid.uuid4())
    files = {'file': open(audio, 'rb')}
    data = {
        'language': 'ms' if source_lang == 'ma' else source_lang,
        'model_name': 'whisper-large-v2-local-cs'
    }

    asr_response = requests.post(ASR_API, files=files, data=data)
    print(asr_response.json())
    if asr_response.status_code == 200:
        transcription = asr_response.json()['text']
    else:
        return "ASR failed", None, None

    translation_prompt = f"Translate the following text from {LANGUAGE_MAP[source_lang]} to {LANGUAGE_MAP[target_lang]}: {transcription}"
    translated_text = inference_via_llm_api(translation_prompt)
    print(f"Translation: {translated_text}")

    # TTS
    tts_params = {
        'language': target_lang,
        'speed': 1.1,
        'speaker': 'MS' if target_lang == 'en' else 'msFemale' if target_lang == 'ma' else 'ta_female1' if target_lang == 'ta' else 'childChinese2',
        'text': translated_text
    }
    
    tts_response = requests.get(TTS_SPEAK_SERVICE, params=tts_params)
    if tts_response.status_code == 200:
        audio_file = tts_response.text.strip()
        audio_url = f"{TTS_WAVE_SERVICE}?file={audio_file}"
        return transcription, translated_text, audio_url
    else:
        return transcription, translated_text, "TTS failed"

def check_password(password):
    return password == DEVELOPER_PASSWORD

def user_interface(audio, source_lang, target_lang):
    _, _, audio_url = transcribe_and_speak(audio, source_lang, target_lang)
    return audio_url

with gr.Blocks() as demo:
    gr.Markdown("# Speech Translation")
    
    with gr.Tab("User Mode"):
        gr.Markdown("Speak into the microphone or upload an audio file. The app will translate and speak it back to you.")
        
        with gr.Row():
            user_audio_input = gr.Audio(sources=["microphone", "upload"], type="filepath")
            user_source_lang = gr.Dropdown(choices=["en", "ma", "ta", "zh"], label="Source Language", value="en")
            user_target_lang = gr.Dropdown(choices=["en", "ma", "ta", "zh"], label="Target Language", value="zh")
        
        with gr.Row():
            user_button = gr.Button("Translate and Speak")
        
        with gr.Row():
            user_audio_output = gr.Audio(label="Translated Speech")
        
        user_button.click(
            fn=user_interface,
            inputs=[user_audio_input, user_source_lang, user_target_lang],
            outputs=[user_audio_output]
        )
    
    with gr.Tab("Developer Mode"):
        password_input = gr.Textbox(type="password", label="Enter Developer Password")
        login_button = gr.Button("Login")
        login_error = gr.Markdown(visible=False)
        
        dev_interface = gr.Column(visible=False)
        
        with dev_interface:
            gr.Markdown("Developer Mode: Transcription, Translation, and TTS")
            
            with gr.Row():
                dev_audio_input = gr.Audio(sources=["microphone", "upload"], type="filepath")
                dev_source_lang = gr.Dropdown(choices=["en", "ma", "ta", "zh"], label="Source Language", value="en")
                dev_target_lang = gr.Dropdown(choices=["en", "ma", "ta", "zh"], label="Target Language", value="zh")
            
            with gr.Row():
                dev_button = gr.Button("Transcribe, Translate, and Speak")
            
            with gr.Row():
                dev_text_output = gr.Textbox(label="Transcription")
            
            with gr.Row():
                dev_translation_output = gr.Textbox(label="Translation")
            
            with gr.Row():
                dev_audio_output = gr.Audio(label="Translated Speech")
            
            dev_button.click(
                fn=transcribe_and_speak,
                inputs=[dev_audio_input, dev_source_lang, dev_target_lang],
                outputs=[dev_text_output, dev_translation_output, dev_audio_output]
            )
        
        def login(password):
            if check_password(password):
                return gr.Column(visible=True), gr.Markdown(visible=False)
            else:
                return gr.Column(visible=False), gr.Markdown("Incorrect password. Please try again.", visible=True)
        
        login_button.click(
            fn=login,
            inputs=[password_input],
            outputs=[dev_interface, login_error]
        )

demo.launch(auth=(os.getenv("DEV_USER"), os.getenv("DEV_PWD")))