stt-server.py

from flask import Flask, Response, request, jsonify
from tempfile import NamedTemporaryFile
from faster_whisper import WhisperModel
import re
import time
import os
from statistics import mean
import logging
from unidecode import unidecode
import multiprocessing

app = Flask(__name__)

models = {}

app.logger.setLevel(logging.INFO)
app.logger.info("Started")

def str2bool(v):
  return v.lower() in ("yes", "true", "on", "y")

@app.route('/api/speech-to-text', methods=['POST'])
def handler():
    app.logger.info("Incoming connection from %s", request.host)
    modelName = request.args.get('model', 'tiny')
    beamSize = int(request.args.get('beamSize', 5))
    language = request.args.get('language', 'ro')
    stripAccents = str2bool(request.args.get('stripAccents', "no"))
    if modelName not in models:
        app.logger.warning("Model %s was not cached", modelName)
        models[modelName] = WhisperModel(modelName, device="cpu", compute_type="int8", cpu_threads=multiprocessing.cpu_count())    
    model = models[modelName]
    data = request.get_data()
    temp = NamedTemporaryFile(suffix='.wav', delete=False)
    app.logger.info("Saving audio stream to %s", temp.name)
    temp.write(data)    
    temp.close()
    start_time = time.perf_counter()
    iterable, _ = model.transcribe(temp.name, language=language, beam_size=beamSize, temperature=0, without_timestamps = True)    
    app.logger.info("Processing audio stream")
    segments = list(iterable)
    end_time = time.perf_counter()
    app.logger.info("Audio stream converted to text in %f seconds", end_time - start_time)
    average = mean(segment.avg_log_prob for segment in segments)
    
    os.remove(temp.name)
    text = " ".join(segment.text for segment in segments)

    clean_text = re.sub(r"[,.;@#?!&$]+\ *", " ", text).strip().lower()
    if stripAccents:
       clean_text = unidecode(clean_text)
    return jsonify(
        text = clean_text,
        original_text = text,
        transcribe_seconds = (end_time - start_time),
        likelihood = 1 + average,
        model = modelName,
        beamSize = beamSize,
        language = language,
        stripAccents = stripAccents
    )