import spaces
import gradio as gr
import io
import os
import re
import torch
import torchaudio
from pathlib import Path
from whisperspeech.pipeline import Pipeline
DEVEL=os.environ.get('DEVEL', False)
title = """
# Welcome to Collabora's WhisperSpeech
WhisperSpeech is an Open Source text-to-speech system built by Collabora and LAION by inverting Whisper.
The model is fully open and you can run it on your local hardware. It's like **Stable Diffusion but for speech**
– both powerful and easily customizable.
[You can contribute to WhisperSpeech on Github.](https://github.com/collabora/WhisperSpeech)
You can also join the discussion on Discord [](https://discord.gg/FANw4rHD5E)
Huge thanks to [Tonic](https://huggingface.co/Tonic) who helped build this Space for WhisperSpeech.
### How to Use It
Write you text in the box, you can use language tags (`` or ``) to create multilingual speech.
Optionally you can upload a speech sample or give it a file URL to clone an existing voice. Check out the
examples at the bottom of the page for inspiration.
"""
footer = """
### How to use it locally
```
pip install -U WhisperSpeech
```
Afterwards:
```
from whisperspeech.pipeline import Pipeline
pipe = Pipeline(torch_compile=True)
pipe.generate_to_file("output.wav", "Hello from WhisperSpeech.")
```
"""
text_examples = [
["This is the first demo of Whisper Speech, a fully open source text-to-speech model trained by Collabora and Lion on the Juwels supercomputer.", None],
["World War II or the Second World War was a global conflict that lasted from 1939 to 1945. The vast majority of the world's countries, including all the great powers, fought as part of two opposing military alliances: the Allies and the Axis.", "https://upload.wikimedia.org/wikipedia/commons/7/75/Winston_Churchill_-_Be_Ye_Men_of_Valour.ogg"],
["To jest pierwszy test wielojęzycznego Whisper Speech , modelu zamieniającego tekst na mowę, który Collabora i Laion nauczyli na superkomputerze Jewels.", None],
[" WhisperSpeech is an Open Source library that helps you convert text to speech. Teraz także po Polsku! I think I just tried saying \"now also in Polish\", don't judge me...", None],
# [" WhisperSpeech is multi-lingual y puede cambiar de idioma मध्य वाक्य में"],
["To jest pierwszy test naszego modelu. Pozdrawiamy serdecznie.", None],
# [" The big difference between Europe et les Etats Unis jest to, że mamy tak wiele języków тут, в Європі"]
]
def parse_multilingual_text(input_text):
pattern = r"(?:<(\w+)>)|([^<]+)"
cur_lang = 'en'
segments = []
for i, (lang, txt) in enumerate(re.findall(pattern, input_text)):
if lang: cur_lang = lang
else: segments.append((cur_lang, f" {txt} ")) # add spaces to give it some time to switch languages
if not segments: return [("en", "")]
return segments
@spaces.GPU(enable_queue=True)
def generate_audio(pipe, segments, speaker, speaker_url, cps=14):
if isinstance(speaker, (str, Path)): speaker = pipe.extract_spk_emb(speaker)
elif speaker_url: speaker = pipe.extract_spk_emb(speaker_url)
else: speaker = pipe.default_speaker
langs, texts = [list(x) for x in zip(*segments)]
print(texts, langs)
stoks = pipe.t2s.generate(texts, cps=cps, lang=langs)
stoks = stoks[stoks!=512]
atoks = pipe.s2a.generate(stoks, speaker.unsqueeze(0))
audio = pipe.vocoder.decode(atoks)
return audio.cpu()
def whisper_speech_demo(multilingual_text, speaker_audio=None, speaker_url="", cps=14):
if len(multilingual_text) == 0:
raise gr.Error("Please enter some text for me to speak!")
segments = parse_multilingual_text(multilingual_text)
audio = generate_audio(pipe, segments, speaker_audio, speaker_url, cps)
return (24000, audio.T.numpy())
# Did not work for me in Safari:
# mp3 = io.BytesIO()
# torchaudio.save(mp3, audio, 24000, format='mp3')
# return mp3.getvalue()
pipe = Pipeline(torch_compile=not DEVEL)
# warmup will come from regenerating the examples
with gr.Blocks() as demo:
gr.Markdown(title)
with gr.Row(equal_height=True):
with gr.Column(scale=2):
text_input = gr.Textbox(label="Enter multilingual text💬📝",
value=text_examples[0][0],
info="You can use `` for English and `` for Polish, see examples below.")
cps = gr.Slider(value=14, minimum=10, maximum=15, step=.25,
label="Tempo (in characters per second)")
with gr.Row(equal_height=True):
speaker_input = gr.Audio(label="Upload or Record Speaker Audio (optional)🌬️💬",
sources=["upload", "microphone"],
type='filepath')
url_input = gr.Textbox(label="alternatively, you can paste in an audio file URL:")
gr.Markdown(" \n ") # fixes the bottom overflow from Audio
generate_button = gr.Button("Try Collabora's WhisperSpeech🌟")
with gr.Column(scale=1):
output_audio = gr.Audio(label="WhisperSpeech says…")
with gr.Column():
gr.Markdown("### Try these examples to get started !🌟🌬️")
gr.Examples(
examples=text_examples,
inputs=[text_input, url_input],
outputs=[output_audio],
fn=whisper_speech_demo,
cache_examples=not DEVEL,
)
generate_button.click(whisper_speech_demo, inputs=[text_input, speaker_input, url_input, cps], outputs=output_audio)
gr.Markdown(footer)
demo.launch(server_port=3000 if DEVEL else None)
