live_api.py

# -*- coding: utf-8 -*-
import numpy as np
import soundfile
import audresample
import text_utils

import re
import subprocess
import markdown
import json
from pathlib import Path
from types import SimpleNamespace
from flask import Flask, request, send_from_directory
from flask_cors import CORS
from audiocraft.builders import AudioGen #, audio_write
NUM_SOUND_GENERATIONS = 1  # they differ a lot and are unnatural to concatenate, prefer lm.n_draw
sound_generator = AudioGen(duration=.74, device='cuda:0').to('cuda:0').eval()

    
# ====STYLE VECTOR====



# AFFECTIVE = True
# VOICE = 'en_UK/apope_low'  #  	en_US/m-ailabs_low#mary_ann

# _dir = '/' if AFFECTIVE else '_v2/'
# precomputed_style_vector = msinference.compute_style(
#     'assets/wavs/style_vector' + _dir + VOICE.replace(
#         '/', '_').replace(
#         '#', '_').replace(
#         'cmu-arctic', 'cmu_arctic').replace(
#         '_low', '') + '.wav')
# print('\n  STYLE VECTOR \n', precomputed_style_vector.shape)


# ==== STYLE VECTOR 

CACHE_DIR = 'flask_cache/'
Path(CACHE_DIR).mkdir(parents=True, exist_ok=True)




def tts_multi_sentence(scene=None):
    if scene is not None and len(scene) >= 4:
        print(f'Processing: {scene} ..')
        # x = sound_generator.generate([scene])[0, :, :].detach().cpu().numpy()
        x = sound_generator.generate(
                                        [scene] * NUM_SOUND_GENERATIONS
                                        ).reshape(1, -1).detach().cpu().numpy() # bs, 11400
        
        x /= np.abs(x).max() + 1e-7
        # is 16kHz - AUdiogen Fs
        x = audresample.resample(x,
            original_rate=16000,
            target_rate=24000)[0, :]
        
            
        #
        print(f'Craft Finished for: {scene}\n\n\n\n____{x.shape}')
    else:
        print(scene, '\nDrop\n')
        x = np.zeros(400)
        
    # # StyleTTS2
    # if ('en_US/' in voice) or ('en_UK/' in voice) or (voice is None):
    #     assert precomputed_style_vector is not None, 'For affective TTS, style vector is needed.'
    #     x = []
    #     for _sentence in text:
    #         x.append(msinference.inference(_sentence,
    #                     precomputed_style_vector,
    #                                 alpha=0.3,
    #                                 beta=0.7,
    #                                 diffusion_steps=7,
    #                                 embedding_scale=1))
    #     x = np.concatenate(x)
        
    #     return overlay(x, sound_background)
    
    return x
    
    




app = Flask(__name__)
cors = CORS(app)


@app.route("/")
def index():
    with open('README.md', 'r') as f:
        return markdown.markdown(f.read())


@app.route("/", methods=['GET', 'POST', 'PUT'])
def serve_wav():
    # https://stackoverflow.com/questions/13522137/in-flask-convert-form-post-
    #                      object-into-a-representation-suitable-for-mongodb
    r = request.form.to_dict(flat=False)
    

    args = SimpleNamespace(
        text=None if r.get('text') is None else r.get('text'),  # string not file?
        scene=r.get('scene')[0]
        )
    # print('\n==RECOMPOSED as \n',request.data,request.form,'\n==')
    




    
    
    x = tts_multi_sentence(args.scene)
    
    OUT_FILE = 'tmp.wav'
    soundfile.write(CACHE_DIR + OUT_FILE, x, 16000)


    
    
    
    # send server's output as default file -> srv_result.xx
    print(f'\n=SERVER saved as {OUT_FILE=}\n')
    response = send_from_directory(CACHE_DIR, path=OUT_FILE)
    response.headers['suffix-file-type'] = OUT_FILE
    return response


if __name__ == "__main__":
    app.run(host="0.0.0.0")