app.py

import spaces
import gradio as gr
import torch
from transformers.models.speecht5.number_normalizer import EnglishNumberNormalizer
from string import punctuation
import re


from parler_tts import ParlerTTSForConditionalGeneration
from transformers import AutoTokenizer, AutoFeatureExtractor, set_seed

device = "cuda:0" if torch.cuda.is_available() else "cpu"


repo_id =  "parler-tts/parler-tts-mini-v1"
repo_id_large = "parler-tts/parler-tts-large-v1"

model = ParlerTTSForConditionalGeneration.from_pretrained(repo_id).to(device)
model_large = ParlerTTSForConditionalGeneration.from_pretrained(repo_id_large).to(device)
tokenizer = AutoTokenizer.from_pretrained(repo_id)
feature_extractor = AutoFeatureExtractor.from_pretrained(repo_id)


SAMPLE_RATE = feature_extractor.sampling_rate
SEED = 42

default_text = "All of the data, pre-processing, training code, and weights are released publicly under a permissive license, enabling the community to build on our work and develop their own powerful models."
default_description = "Laura's voice is monotone yet slightly fast in delivery, with a very close recording that almost has no background noise."
examples = [
    [
        "This version introduces speaker consistency across generations, characterized by their name. For example, Jon, Lea, Gary, Jenna, Mike and Laura.",
        "Gary's voice is monotone yet slightly fast in delivery, with a very close recording that has no background noise.",
        None,
    ],
    [
        '''There's 34 speakers. To take advantage of this, simply adapt your text description to specify which speaker to use: "Mike speaks animatedly...".''',
        "Gary speaks slightly animatedly and slightly slowly in delivery, with a very close recording that has no background noise.",
        None
    ],
    [
        "'This is the best time of my life, Bartley,' she said happily.",
        "A female speaker delivers a slightly expressive and animated speech with a moderate speed. The recording features a low-pitch voice and slight background noise, creating a close-sounding audio experience.",
        None,
    ],
    [
        "Montrose also, after having experienced still more variety of good and bad fortune, threw down his arms, and retired out of the kingdom.",
        "A man voice speaks slightly slowly with very noisy background, carrying a low-pitch tone and displaying a touch of expressiveness and animation. The sound is very distant, adding an air of intrigue.",
        None
    ],
    [
        "Once upon a time, in the depth of winter, when the flakes of snow fell like feathers from the clouds, a queen sat sewing at her pal-ace window, which had a carved frame of black wood.",
        "In a very poor recording quality, a female speaker delivers her slightly expressive and animated words with a fast pace. There's high level of background noise and a very distant-sounding reverberation. Her voice is slightly higher pitched than average.",
        None,
    ],
]

number_normalizer = EnglishNumberNormalizer()

def preprocess(text):
    text = number_normalizer(text).strip()
    text = text.replace("-", " ")
    if text[-1] not in punctuation:
        text = f"{text}."
    
    abbreviations_pattern = r'\b[A-Z][A-Z\.]+\b'
    
    def separate_abb(chunk):
        chunk = chunk.replace(".","")
        print(chunk)
        return " ".join(chunk)
    
    abbreviations = re.findall(abbreviations_pattern, text)
    for abv in abbreviations:
        if abv in text:
            text = text.replace(abv, separate_abb(abv))
    return text

@spaces.GPU
def gen_tts(text, description, use_large=False):
    inputs = tokenizer(description.strip(), return_tensors="pt").to(device)
    prompt = tokenizer(preprocess(text), return_tensors="pt").to(device)

    set_seed(SEED)
    if use_large:
        generation = model_large.generate(
            input_ids=inputs.input_ids, prompt_input_ids=prompt.input_ids, attention_mask=inputs.attention_mask, prompt_attention_mask=prompt.attention_mask, do_sample=True, temperature=1.0
        )      
    else:
        generation = model.generate(
            input_ids=inputs.input_ids, prompt_input_ids=prompt.input_ids, attention_mask=inputs.attention_mask, prompt_attention_mask=prompt.attention_mask, do_sample=True, temperature=1.0
        )
    audio_arr = generation.cpu().numpy().squeeze()

    return SAMPLE_RATE, audio_arr


css = """
        #share-btn-container {
            display: flex;
            padding-left: 0.5rem !important;
            padding-right: 0.5rem !important;
            background-color: #000000;
            justify-content: center;
            align-items: center;
            border-radius: 9999px !important; 
            width: 13rem;
            margin-top: 10px;
            margin-left: auto;
            flex: unset !important;
        }
        #share-btn {
            all: initial;
            color: #ffffff;
            font-weight: 600;
            cursor: pointer;
            font-family: 'IBM Plex Sans', sans-serif;
            margin-left: 0.5rem !important;
            padding-top: 0.25rem !important;
            padding-bottom: 0.25rem !important;
            right:0;
        }
        #share-btn * {
            all: unset !important;
        }
        #share-btn-container div:nth-child(-n+2){
            width: auto !important;
            min-height: 0px !important;
        }
        #share-btn-container .wrap {
            display: none !important;
        }
"""
with gr.Blocks(css=css) as block:
    gr.HTML(
        """
            <div style="text-align: center; max-width: 700px; margin: 0 auto;">
              <div
                style="
                  display: inline-flex; align-items: center; gap: 0.8rem; font-size: 1.75rem;
                "
              >
                <h1 style="font-weight: 900; margin-bottom: 7px; line-height: normal;">
                  Parler-TTS 🗣️
                </h1>
              </div>
            </div>
        """
    )
    gr.HTML(
        f"""
        <p><a href="https://github.com/huggingface/parler-tts"> Parler-TTS</a> is a training and inference library for
        high-fidelity text-to-speech (TTS) models.</p> 
        <p>The models demonstrated here, Parler-TTS <a href="https://huggingface.co/parler-tts/parler-tts-mini-v1">Mini v1</a> and <a href="https://huggingface.co/parler-tts/parler-tts-large-v1">Large v1</a>, 
        are trained using 45k hours of narrated English audiobooks. It generates high-quality speech 
        with features that can be controlled using a simple text prompt (e.g. gender, background noise, speaking rate, pitch and reverberation).</p>
        <p>By default, Parler-TTS generates 🎲 random voice. To ensure 🎯 <b> speaker consistency </b> across generations, these checkpoints were also trained on 34 speakers, characterized by name (e.g. Jon, Lea, Gary, Jenna, Mike, Laura). Learn more about this <a href="https://github.com/huggingface/parler-tts/blob/main/INFERENCE.md#speaker-consistency"> here </a>.</p>
        
        <p>To take advantage of this, simply adapt your text description to specify which speaker to use: `Jon's voice is monotone...`</p>
        """
    )
    with gr.Row():
        with gr.Column():
            input_text = gr.Textbox(label="Input Text", lines=2, value=default_text, elem_id="input_text")
            description = gr.Textbox(label="Description", lines=2, value=default_description, elem_id="input_description")
            use_large = gr.Checkbox(value=False, label="Use Large checkpoint", info="Generate with Parler-TTS Large v1 instead of Mini v1 - Better but way slower.")
            run_button = gr.Button("Generate Audio", variant="primary")
        with gr.Column():
            audio_out = gr.Audio(label="Parler-TTS generation", type="numpy", elem_id="audio_out")

    inputs = [input_text, description, use_large]
    outputs = [audio_out]
    run_button.click(fn=gen_tts, inputs=inputs, outputs=outputs, queue=True)
    gr.Examples(examples=examples, fn=gen_tts, inputs=inputs, outputs=outputs, cache_examples=True)
    gr.HTML(
        """
        <p>Tips for ensuring good generation:
        <ul>
            <li>Include the term "very clear audio" to generate the highest quality audio, and "very noisy audio" for high levels of background noise</li>
            <li>Punctuation can be used to control the prosody of the generations, e.g. use commas to add small breaks in speech</li>
            <li>The remaining speech features (gender, speaking rate, pitch and reverberation) can be controlled directly through the prompt</li>
        </ul>
        </p>
        <p>Parler-TTS can be much faster. We give some tips on how to generate much more quickly in this <a href="https://github.com/huggingface/parler-tts/blob/main/INFERENCE.md"> inference guide</a>. Think SDPA, torch.compile, batching and streaming!</p>
        
        <p>If you want to find out more about how this model was trained and even fine-tune it yourself, check-out the 
        <a href="https://github.com/huggingface/parler-tts"> Parler-TTS</a> repository on GitHub.</p>
        
        <p>The Parler-TTS codebase and its associated checkpoints are licensed under <a href='https://github.com/huggingface/parler-tts?tab=Apache-2.0-1-ov-file#readme'> Apache 2.0</a>.</p>
        """
    )


block.queue()
block.launch(share=True)