app.py

import os
import re
from num2words import num2words
import gradio as gr
import torch
import torchaudio
from data.tokenizer import (
    AudioTokenizer,
    TextTokenizer,
)
from edit_utils_zh import parse_edit_zh
from edit_utils_en import parse_edit_en
from edit_utils_zh import parse_tts_zh
from edit_utils_en import parse_tts_en
from inference_scale import inference_one_sample
import librosa
import soundfile as sf
from models import ssr
import io
import numpy as np
import random
import uuid
import opencc
import spaces
import nltk
nltk.download('punkt')

DEMO_PATH = os.getenv("DEMO_PATH", "./demo")
TMP_PATH = os.getenv("TMP_PATH", "./demo/temp")
MODELS_PATH = os.getenv("MODELS_PATH", "./pretrained_models")
device = "cuda" if torch.cuda.is_available() else "cpu"
whisper_model, align_model, ssrspeech_model = None, None, None

def get_random_string():
    return "".join(str(uuid.uuid4()).split("-"))

def traditional_to_simplified(segments):
    converter = opencc.OpenCC('t2s') 
    seg_num = len(segments)
    for i in range(seg_num):
        words = segments[i]['words']
        for j in range(len(words)):
            segments[i]['words'][j]['word'] = converter.convert(segments[i]['words'][j]['word'])
        segments[i]['text'] = converter.convert(segments[i]['text'])
    return segments

@spaces.GPU
def seed_everything(seed):
    if seed != -1:
        os.environ['PYTHONHASHSEED'] = str(seed)
        random.seed(seed)
        np.random.seed(seed)
        torch.manual_seed(seed)
        torch.cuda.manual_seed(seed)
        torch.backends.cudnn.benchmark = False
        torch.backends.cudnn.deterministic = True

def get_mask_interval(transcribe_state, word_span):
    print(transcribe_state)
    seg_num = len(transcribe_state['segments'])
    data = []
    for i in range(seg_num):
      words = transcribe_state['segments'][i]['words']
      for item in words:
        data.append([item['start'], item['end'], item['word']])

    s, e = word_span[0], word_span[1]
    assert s <= e, f"s:{s}, e:{e}"
    assert s >= 0, f"s:{s}"
    assert e <= len(data), f"e:{e}"
    if e == 0: # start
        start = 0.
        end = float(data[0][0])
    elif s == len(data): # end
        start = float(data[-1][1])
        end = float(data[-1][1]) # don't know the end yet
    elif s == e: # insert
        start = float(data[s-1][1])
        end = float(data[s][0])
    else:
        start = float(data[s-1][1]) if s > 0 else float(data[s][0])
        end = float(data[e][0]) if e < len(data) else float(data[-1][1])

    return (start, end)
    
@spaces.GPU
class WhisperxAlignModel:
    def __init__(self, language):
        from whisperx import load_align_model
        self.model, self.metadata = load_align_model(language_code=language, device=device)

    def align(self, segments, audio_path):
        from whisperx import align, load_audio
        audio = load_audio(audio_path)
        return align(segments, self.model, self.metadata, audio, device, return_char_alignments=False)["segments"]

@spaces.GPU
class WhisperModel:
    def __init__(self, model_name, language):
        from whisper import load_model
        self.model = load_model(model_name, device, language=language)

        from whisper.tokenizer import get_tokenizer
        tokenizer = get_tokenizer(multilingual=False, language=language)
        self.supress_tokens = [-1] + [
            i
            for i in range(tokenizer.eot)
            if all(c in "0123456789" for c in tokenizer.decode([i]).removeprefix(" "))
        ]

    def transcribe(self, audio_path):
        return self.model.transcribe(audio_path, suppress_tokens=self.supress_tokens, word_timestamps=True)["segments"]

@spaces.GPU
class WhisperxModel:
    def __init__(self, model_name, align_model, language):
        from whisperx import load_model
        self.model = load_model(model_name, device, asr_options={"suppress_numerals": True, "max_new_tokens": None, "clip_timestamps": None, "hallucination_silence_threshold": None}, language=language)
        self.align_model = align_model

    def transcribe(self, audio_path):
        segments = self.model.transcribe(audio_path, batch_size=8)["segments"]
        for segment in segments:
            segment['text'] = replace_numbers_with_words(segment['text'])
        return self.align_model.align(segments, audio_path)

@spaces.GPU
def load_models(whisper_backend_name, ssrspeech_model_name):
    global transcribe_model, align_model, ssrspeech_model
    
    alignment_model_name = "whisperX"
    if ssrspeech_model_name == "English":
        ssrspeech_model_name = "English"
        text_tokenizer = TextTokenizer(backend="espeak")
        language = "en"
        whisper_model_name = "base.en"
        
    elif ssrspeech_model_name == "Mandarin":
        ssrspeech_model_name = "Mandarin"
        text_tokenizer = TextTokenizer(backend="espeak", language='cmn')
        language = "zh"
        whisper_model_name = "base"

    if alignment_model_name is not None:
        align_model = WhisperxAlignModel(language)

    if whisper_model_name is not None:
        if whisper_backend_name == "whisper":
            transcribe_model = WhisperModel(whisper_model_name, language)
        else:
            if align_model is None:
                raise gr.Error("Align model required for whisperx backend")
            transcribe_model = WhisperxModel(whisper_model_name, align_model, language)

    ssrspeech_fn = f"{MODELS_PATH}/{ssrspeech_model_name}.pth"
    if not os.path.exists(ssrspeech_fn):
        os.system(f"wget https://huggingface.co/westbrook/SSR-Speech-{ssrspeech_model_name}/resolve/main/{ssrspeech_model_name}.pth -O " + ssrspeech_fn)
    print(transcribe_model, align_model)
    ckpt = torch.load(ssrspeech_fn)
    model = ssr.SSR_Speech(ckpt["config"])
    model.load_state_dict(ckpt["model"])
    config = model.args
    phn2num = ckpt["phn2num"]
    model.to(device)

    encodec_fn = f"{MODELS_PATH}/wmencodec.th"
    if not os.path.exists(encodec_fn):
        os.system(f"wget https://huggingface.co/westbrook/SSR-Speech-English/resolve/main/wmencodec.th -O " + encodec_fn)

    ssrspeech_model = {
        "config": config,
        "phn2num": phn2num,
        "model": model,
        "text_tokenizer": text_tokenizer,
        "audio_tokenizer": AudioTokenizer(signature=encodec_fn)
    }
    success_message = "<span style='color:green;'>Success: Models loading completed successfully!</span>"

    return [
        gr.Accordion(),
        success_message
    ]


def get_transcribe_state(segments):
    transcript = " ".join([segment["text"] for segment in segments])
    transcript = transcript[1:] if transcript[0] == " " else transcript
    return {
        "segments": segments,
        "transcript": transcript,
    }

@spaces.GPU
def transcribe(audio_path):
    global transcribe_model
    
    if transcribe_model is None:
        raise gr.Error("Transcription model not loaded")

    segments = transcribe_model.transcribe(audio_path)
    state = get_transcribe_state(segments)
    success_message = "<span style='color:green;'>Success: Transcribe completed successfully!</span>"

    return [
        state["transcript"], state['segments'],
        state, success_message
    ]


@spaces.GPU
def align(segments, audio_path):
    global align_model
    if align_model is None:
        raise gr.Error("Align model not loaded")
    
    segments = align_model.align(segments, audio_path)
    state = get_transcribe_state(segments)

    return state


def get_output_audio(audio_tensors, codec_audio_sr):
    result = torch.cat(audio_tensors, 1)
    buffer = io.BytesIO()
    torchaudio.save(buffer, result, int(codec_audio_sr), format="wav")
    buffer.seek(0)
    return buffer.read()

def replace_numbers_with_words(sentence):
    sentence = re.sub(r'(\d+)', r' \1 ', sentence) # add spaces around numbers
    def replace_with_words(match):
        num = match.group(0)
        try:
            return num2words(num) # Convert numbers to words
        except:
            return num # In case num2words fails (unlikely with digits but just to be safe)
    return re.sub(r'\b\d+\b', replace_with_words, sentence) # Regular expression that matches numbers

@spaces.GPU
def run(seed, sub_amount, ssrspeech_model_choice, codec_audio_sr, codec_sr, top_k, top_p, temperature,
        stop_repetition, kvcache, silence_tokens, aug_text, cfg_coef, prompt_length,
        audio_path, original_transcript, transcript, mode):
    
    global transcribe_model, align_model, ssrspeech_model
    aug_text = True if aug_text == 1 else False
    if ssrspeech_model is None:
        raise gr.Error("ssrspeech model not loaded")
    
    seed_everything(seed)
    
    if ssrspeech_model_choice == "English":
        language = "en"
    elif ssrspeech_model_choice == "Mandarin":
        language = "zh"

    # resample audio
    audio, _ = librosa.load(audio_path, sr=16000)
    sf.write(audio_path, audio, 16000)
    
    # text normalization
    target_transcript = replace_numbers_with_words(transcript).replace("  ", " ").replace("  ", " ").replace("\n", " ")
    orig_transcript = replace_numbers_with_words(original_transcript).replace("  ", " ").replace("  ", " ").replace("\n", " ")

    orig_transcript, segments = transcribe(audio_path)
    if language == 'zh':
        converter = opencc.OpenCC('t2s')
        orig_transcript = converter.convert(orig_transcript)
        transcribe_state = align(traditional_to_simplified(segments), audio_path)
        transcribe_state['segments'] = traditional_to_simplified(transcribe_state['segments'])
    elif language == 'en':
        orig_transcript = orig_transcript.lower()
        target_transcript = target_transcript.lower()
        transcribe_state = align(segments, audio_path)
    print(orig_transcript)
    print(target_transcript)

    if mode == "TTS":
        info = torchaudio.info(audio_path)
        duration = info.num_frames / info.sample_rate
        cut_length = duration
        # Cut long audio for tts
        if duration > prompt_length:
            seg_num = len(transcribe_state['segments'])
            for i in range(seg_num):
                words = transcribe_state['segments'][i]['words']
                for item in words:
                    if item['end'] >= prompt_length:
                        cut_length = min(item['end'], cut_length)

        audio, _ = librosa.load(audio_path, sr=16000, duration=cut_length)
        sf.write(audio_path, audio, 16000)
        orig_transcript, segments = transcribe(audio_path)
        
        if language == 'zh':
            converter = opencc.OpenCC('t2s')
            orig_transcript = converter.convert(orig_transcript)
            transcribe_state = align(traditional_to_simplified(segments), audio_path)
            transcribe_state['segments'] = traditional_to_simplified(transcribe_state['segments'])
        elif language == 'en':
            orig_transcript = orig_transcript.lower()
            target_transcript = target_transcript.lower()
            transcribe_state = align(segments, audio_path)
        print(orig_transcript)
        target_transcript_copy = target_transcript # for tts cut out
        if language == 'en':
            target_transcript_copy = target_transcript_copy.split(' ')[0]
        elif language == 'zh':
            target_transcript_copy = target_transcript_copy[0]
        target_transcript = orig_transcript + ' ' + target_transcript if language == 'en' else orig_transcript + target_transcript
        print(target_transcript)

    if mode == "Edit":
        operations, orig_spans = parse_edit_en(orig_transcript, target_transcript) if language == 'en' else parse_edit_zh(orig_transcript, target_transcript)
        print(operations)
        print("orig_spans: ", orig_spans)
        
        if len(orig_spans) > 3:
            raise gr.Error("Current model only supports maximum 3 editings")
            
        starting_intervals = []
        ending_intervals = []
        for orig_span in orig_spans:
            start, end = get_mask_interval(transcribe_state, orig_span)
            starting_intervals.append(start)
            ending_intervals.append(end)
    
        print("intervals: ", starting_intervals, ending_intervals)
    
        info = torchaudio.info(audio_path)
        audio_dur = info.num_frames / info.sample_rate
        
        def combine_spans(spans, threshold=0.2):
            spans.sort(key=lambda x: x[0])
            combined_spans = []
            current_span = spans[0]
    
            for i in range(1, len(spans)):
                next_span = spans[i]
                if current_span[1] >= next_span[0] - threshold:
                    current_span[1] = max(current_span[1], next_span[1])
                else:
                    combined_spans.append(current_span)
                    current_span = next_span
            combined_spans.append(current_span)
            return combined_spans
        
        morphed_span = [[max(start - sub_amount, 0), min(end + sub_amount, audio_dur)]
                        for start, end in zip(starting_intervals, ending_intervals)] # in seconds
        morphed_span = combine_spans(morphed_span, threshold=0.2)
        print("morphed_spans: ", morphed_span)
        mask_interval = [[round(span[0]*codec_sr), round(span[1]*codec_sr)] for span in morphed_span]
        mask_interval = torch.LongTensor(mask_interval) # [M,2], M==1 for now
    else:
        info = torchaudio.info(audio_path)
        audio_dur = info.num_frames / info.sample_rate
        
        morphed_span = [(audio_dur, audio_dur)] # in seconds
        mask_interval = [[round(span[0]*codec_sr), round(span[1]*codec_sr)] for span in morphed_span]
        mask_interval = torch.LongTensor(mask_interval) # [M,2], M==1 for now
        print("mask_interval: ", mask_interval)

    decode_config = {'top_k': top_k, 'top_p': top_p, 'temperature': temperature, 'stop_repetition': stop_repetition, 'kvcache': kvcache, "codec_audio_sr": codec_audio_sr, "codec_sr": codec_sr}
    
    tts = True if mode == "TTS" else False
    new_audio = inference_one_sample(
        ssrspeech_model["model"],
        ssrspeech_model["config"],
        ssrspeech_model["phn2num"],
        ssrspeech_model["text_tokenizer"], 
        ssrspeech_model["audio_tokenizer"],
        audio_path, orig_transcript, target_transcript, mask_interval,
        cfg_coef, aug_text, False, True, tts,
        device, decode_config
    )
    audio_tensors = []
    # save segments for comparison
    new_audio = new_audio[0].cpu()
    torchaudio.save(audio_path, new_audio, codec_audio_sr)
    if tts: # remove the start parts
        new_transcript, new_segments = transcribe(audio_path)
        if language == 'zh':
            transcribe_state = align(traditional_to_simplified(new_segments), audio_path)
            transcribe_state['segments'] = traditional_to_simplified(transcribe_state['segments'])
            tmp1 = transcribe_state['segments'][0]['words'][0]['word']
            tmp2 = target_transcript_copy
        elif language == 'en':
            transcribe_state = align(new_segments, audio_path)
            tmp1 = transcribe_state['segments'][0]['words'][0]['word'].lower()
            tmp2 = target_transcript_copy.lower()
        if tmp1 == tmp2:
            offset = transcribe_state['segments'][0]['words'][0]['start']
        else:
            offset = transcribe_state['segments'][0]['words'][1]['start']
        
        new_audio, _ = torchaudio.load(audio_path, frame_offset=int(offset*codec_audio_sr))
    audio_tensors.append(new_audio)
    output_audio = get_output_audio(audio_tensors, codec_audio_sr)
    
    success_message = "<span style='color:green;'>Success: Inference successfully!</span>"
    return output_audio, success_message


demo_original_transcript = "Gwynplaine had, besides, for his work and for his feats of strength, round his neck and over his shoulders, an esclavine of leather."

demo_text = {
    "TTS": {
        "regular": "Gwynplaine had, besides, for his work and for his feats of strength, I cannot believe that the same model can also do text to speech synthesis too!"
    },
    "Edit": {
        "regular": "Gwynplaine had, besides, for his work and for his feats of strength, take over the stage for half an hour, an esclavine of leather."
    },
}

all_demo_texts = {vv for k, v in demo_text.items() for kk, vv in v.items()}

demo_words = ['0.069 Gwynplain 0.611', '0.671 had, 0.912', '0.952 besides, 1.414', '1.494 for 1.634', '1.695 his 1.835', '1.915 work 2.136', '2.196 and 2.297', '2.337 for 2.517', '2.557 his 2.678', '2.758 feats 3.019', '3.079 of 3.139', '3.2 strength, 3.561', '4.022 round 4.263', '4.303 his 4.444', '4.524 neck 4.705', '4.745 and 4.825', '4.905 over 5.086', '5.146 his 5.266', '5.307 shoulders, 5.768', '6.23 an 6.33', '6.531 esclavine 7.133', '7.213 of 7.293', '7.353 leather. 7.614']

demo_words_info = [{'word': 'Gwynplain', 'start': 0.069, 'end': 0.611, 'score': 0.833}, {'word': 'had,', 'start': 0.671, 'end': 0.912, 'score': 0.879}, {'word': 'besides,', 'start': 0.952, 'end': 1.414, 'score': 0.863}, {'word': 'for', 'start': 1.494, 'end': 1.634, 'score': 0.89}, {'word': 'his', 'start': 1.695, 'end': 1.835, 'score': 0.669}, {'word': 'work', 'start': 1.915, 'end': 2.136, 'score': 0.916}, {'word': 'and', 'start': 2.196, 'end': 2.297, 'score': 0.766}, {'word': 'for', 'start': 2.337, 'end': 2.517, 'score': 0.808}, {'word': 'his', 'start': 2.557, 'end': 2.678, 'score': 0.786}, {'word': 'feats', 'start': 2.758, 'end': 3.019, 'score': 0.97}, {'word': 'of', 'start': 3.079, 'end': 3.139, 'score': 0.752}, {'word': 'strength,', 'start': 3.2, 'end': 3.561, 'score': 0.742}, {'word': 'round', 'start': 4.022, 'end': 4.263, 'score': 0.916}, {'word': 'his', 'start': 4.303, 'end': 4.444, 'score': 0.666}, {'word': 'neck', 'start': 4.524, 'end': 4.705, 'score': 0.908}, {'word': 'and', 'start': 4.745, 'end': 4.825, 'score': 0.882}, {'word': 'over', 'start': 4.905, 'end': 5.086, 'score': 0.847}, {'word': 'his', 'start': 5.146, 'end': 5.266, 'score': 0.791}, {'word': 'shoulders,', 'start': 5.307, 'end': 5.768, 'score': 0.729}, {'word': 'an', 'start': 6.23, 'end': 6.33, 'score': 0.854}, {'word': 'esclavine', 'start': 6.531, 'end': 7.133, 'score': 0.803}, {'word': 'of', 'start': 7.213, 'end': 7.293, 'score': 0.772}, {'word': 'leather.', 'start': 7.353, 'end': 7.614, 'score': 0.896}]


def get_app():
    with gr.Blocks() as app:
        with gr.Row():
            with gr.Column(scale=2):
                load_models_btn = gr.Button(value="Load models")
            with gr.Column(scale=5):
                with gr.Accordion("Select models", open=False) as models_selector:
                    with gr.Row():
                        ssrspeech_model_choice = gr.Radio(label="ssrspeech model", value="English",
                                                        choices=["English", "Mandarin"])
                        whisper_backend_choice = gr.Radio(label="Whisper backend", value="whisperX", choices=["whisperX", "whisper"])

        with gr.Row():
            with gr.Column(scale=2):
                input_audio = gr.Audio(value=f"{DEMO_PATH}/5895_34622_000026_000002.wav", label="Input Audio", type="filepath", interactive=True)
                with gr.Group():
                    original_transcript = gr.Textbox(label="Original transcript", lines=5, value=demo_original_transcript,
                                                    info="Use whisperx model to get the transcript.")
                    transcribe_btn = gr.Button(value="Transcribe")

            with gr.Column(scale=3):
                with gr.Group():
                    transcript = gr.Textbox(label="Text", lines=7, value=demo_text["TTS"]["regular"])

                    with gr.Row():
                        mode = gr.Radio(label="Mode", choices=["Edit", "TTS"], value="Edit")

                    run_btn = gr.Button(value="Run")

            with gr.Column(scale=2):
                output_audio = gr.Audio(label="Output Audio")
                with gr.Accordion("Inference transcript", open=False):
                    inference_transcript = gr.Textbox(label="Inference transcript", lines=5, interactive=False,
                                                    info="Inference was performed on this transcript.")

        with gr.Row():
            with gr.Accordion("Generation Parameters - change these if you are unhappy with the generation", open=False):
                stop_repetition = gr.Radio(label="stop_repetition", choices=[-1, 1, 2, 3, 4], value=2,
                                        info="if there are long silence in the generated audio, reduce the stop_repetition to 2 or 1. -1 = disabled")
                seed = gr.Number(label="seed", value=-1, precision=0, info="random seeds always works :)")
                kvcache = gr.Radio(label="kvcache", choices=[0, 1], value=1,
                                    info="set to 0 to use less VRAM, but with slower inference")
                aug_text = gr.Radio(label="aug_text", choices=[0, 1], value=1,
                                    info="set to 1 to use cfg")
                cfg_coef = gr.Number(label="cfg_coef", value=1.5,
                                    info="cfg guidance scale, 1.5 is a good value, change if you don't like the results")
                prompt_length = gr.Number(label="prompt_length", value=3,
                                    info="used for tts prompt, will automatically cut the prompt audio to this length")
                sub_amount = gr.Number(label="sub_amount", value=0.12, info="margin to the left and right of the editing segment, change if you don't like the results")
                top_p = gr.Number(label="top_p", value=0.8, info="0.9 is a good value, 0.8 is also good")
                temperature = gr.Number(label="temperature", value=1, info="haven't try other values, do not change")
                top_k = gr.Number(label="top_k", value=0, info="0 means we don't use topk sampling, because we use topp sampling")
                codec_audio_sr = gr.Number(label="codec_audio_sr", value=16000, info='encodec specific, do not change')
                codec_sr = gr.Number(label="codec_sr", value=50, info='encodec specific, do not change')
                silence_tokens = gr.Textbox(label="silence tokens", value="[1388,1898,131]", info="encodec specific, do not change")

        success_output = gr.HTML()

        load_models_btn.click(fn=load_models,
                            inputs=[whisper_backend_choice, ssrspeech_model_choice],
                            outputs=[models_selector, success_output])

        semgents = gr.State() # not used
        transcribe_btn.click(fn=transcribe,
                            inputs=[input_audio],
                            outputs=[original_transcript, semgents, success_output])
        
        run_btn.click(fn=run,
                    inputs=[
                        seed, sub_amount, ssrspeech_model_choice,
                        codec_audio_sr, codec_sr,
                        top_k, top_p, temperature, stop_repetition, kvcache, silence_tokens, 
                        aug_text, cfg_coef, prompt_length, 
                        input_audio, original_transcript, transcript,
                        mode
                    ],
                    outputs=[output_audio, success_output])

    return app


if __name__ == "__main__":
    import argparse

    parser = argparse.ArgumentParser(description="Ssrspeech gradio app.")
    
    parser.add_argument("--demo-path", default="./demo", help="Path to demo directory")
    parser.add_argument("--tmp-path", default="./demo/temp", help="Path to tmp directory")
    parser.add_argument("--models-path", default="./pretrained_models", help="Path to ssrspeech models directory")
    parser.add_argument("--port", default=7860, type=int, help="App port")
    parser.add_argument("--share", action="store_true", help="Launch with public url")

    os.environ["USER"] = os.getenv("USER", "user")
    args = parser.parse_args()
    DEMO_PATH = args.demo_path
    TMP_PATH = args.tmp_path
    MODELS_PATH = args.models_path

    app = get_app()
    app.queue().launch(share=args.share, server_port=args.port)