piper-turkish-tts

# pip install piper-tts huggingface_hub pydub nltk imageio-ffmpeg

import imageio_ffmpeg
from pydub import AudioSegment
AudioSegment.converter = imageio_ffmpeg.get_ffmpeg_exe()

import time
from huggingface_hub import hf_hub_download
from pydub import AudioSegment
from piper import PiperVoice, SynthesisConfig
import nltk
nltk.download('punkt')
nltk.download('punkt_tab')

from nltk.tokenize import sent_tokenize

MODEL_PATH = hf_hub_download(
    repo_id="99eren99/piper-turkish-tts", 
    filename="model.onnx"
)

CONFIG_PATH = hf_hub_download(
    repo_id="99eren99/piper-turkish-tts", 
    filename="config.json"
)

MODEL = PiperVoice.load(MODEL_PATH,config_path=CONFIG_PATH, use_cuda=False)

# ==============================
# TTS FUNCTION
# ==============================
def generate(text, save_path, volume=1, length_scale=1.2, noise_scale=0.4, noise_w_scale=0.3,
            silence_ms_duration_between_setences=400, normalize_audio=False):
    if not text or not text.strip():
        return None

    text = text.replace("...",". ").replace(",",", ")

    syn_config = SynthesisConfig(
        volume=volume,
        length_scale=length_scale,
        noise_scale=noise_scale,
        noise_w_scale=noise_w_scale,
        normalize_audio=normalize_audio,
    )

    t0 = time.time()
    first = True

    combined = AudioSegment.silent(duration=0)

    fade_ms = 5  
    
    silence = AudioSegment.silent(duration=silence_ms_duration_between_setences)
    combined += silence
    
    for sentence in sent_tokenize(text, language="turkish"):
        for audio_chunk in MODEL.synthesize(sentence, syn_config=syn_config):
    
            if first:
                time_to_audio = "Time To First Bytes(seconds): " + str(time.time() - t0)
                first = False
    
            segment = AudioSegment(
                data=audio_chunk.audio_int16_bytes,
                sample_width=2,
                frame_rate=22050,
                channels=1
            )

            segment = segment.fade_in(fade_ms).fade_out(fade_ms)
            combined += segment + silence

    real_time_factor = "Real Time Factor: " + str((time.time() - t0) / combined.duration_seconds)
    time_to_generate = "Total Time To Generate(seconds): "+str(time.time() - t0)
    metrics = "\n".join([real_time_factor,time_to_audio,time_to_generate])
    print(metrics)
    combined.export(save_path, format="mp3", bitrate="320k")

generate("fanfinfon","out.mp3")