Manmay's picture
added streaming faster inference with hifidecoder
840a99d
import os
import torch
import gradio as gr
import torchaudio
import time
from datetime import datetime
from tortoise.api import TextToSpeech
from tortoise.utils.text import split_and_recombine_text
from tortoise.utils.audio import load_audio, load_voice, load_voices
VOICE_OPTIONS = [
"angie",
"deniro",
"freeman",
"halle",
"lj",
"myself",
"pat2",
"snakes",
"tom",
"daws",
"dreams",
"grace",
"lescault",
"weaver",
"applejack",
"daniel",
"emma",
"geralt",
"jlaw",
"mol",
"pat",
"rainbow",
"tim_reynolds",
"atkins",
"dortice",
"empire",
"kennard",
"mouse",
"william",
"jane_eyre",
"random", # special option for random voice
]
def inference(
text,
script,
voice,
voice_b,
seed,
split_by_newline,
):
if text is None or text.strip() == "":
with open(script.name) as f:
text = f.read()
if text.strip() == "":
raise gr.Error("Please provide either text or script file with content.")
if split_by_newline == "Yes":
texts = list(filter(lambda x: x.strip() != "", text.split("\n")))
else:
texts = split_and_recombine_text(text)
voices = [voice]
if voice_b != "disabled":
voices.append(voice_b)
if len(voices) == 1:
voice_samples, conditioning_latents = load_voice(voice)
else:
voice_samples, conditioning_latents = load_voices(voices)
start_time = time.time()
# all_parts = []
for j, text in enumerate(texts):
for audio_frame in tts.tts_with_preset(
text,
voice_samples=voice_samples,
conditioning_latents=conditioning_latents,
preset="ultra_fast",
k=1
):
# print("Time taken: ", time.time() - start_time)
# all_parts.append(audio_frame)
yield (24000, audio_frame.cpu().detach().numpy())
# wav = torch.cat(all_parts, dim=0).unsqueeze(0)
# print(wav.shape)
# torchaudio.save("output.wav", wav.cpu(), 24000)
# yield (None, gr.make_waveform(audio="output.wav",))
def main():
title = "Tortoise TTS 🐢"
description = """
A text-to-speech system which powers lot of organizations in Speech synthesis domain.
<br/>
a model with strong multi-voice capabilities, highly realistic prosody and intonation.
<br/>
for faster inference, use the 'ultra_fast' preset and duplicate space if you don't want to wait in a queue.
<br/>
"""
text = gr.Textbox(
lines=4,
label="Text (Provide either text, or upload a newline separated text file below):",
)
script = gr.File(label="Upload a text file")
voice = gr.Dropdown(
VOICE_OPTIONS, value="jane_eyre", label="Select voice:", type="value"
)
voice_b = gr.Dropdown(
VOICE_OPTIONS,
value="disabled",
label="(Optional) Select second voice:",
type="value",
)
split_by_newline = gr.Radio(
["Yes", "No"],
label="Split by newline (If [No], it will automatically try to find relevant splits):",
type="value",
value="No",
)
output_audio = gr.Audio(label="streaming audio:", streaming=True, autoplay=True)
# download_audio = gr.Audio(label="dowanload audio:")
interface = gr.Interface(
fn=inference,
inputs=[
text,
script,
voice,
voice_b,
split_by_newline,
],
title=title,
description=description,
outputs=[output_audio],
)
interface.queue().launch()
if __name__ == "__main__":
tts = TextToSpeech(kv_cache=True, use_deepspeed=True, half=True)
with open("Tortoise_TTS_Runs_Scripts.log", "a") as f:
f.write(
f"\n\n-------------------------Tortoise TTS Scripts Logs, {datetime.now()}-------------------------\n"
)
main()