import os import subprocess import uuid import time import torch import torchaudio # langid is used to detect language for longer text # Most users expect text to be their own language, there is checkbox to disable it import langid import csv from io import StringIO import datetime import re import gradio as gr from TTS.tts.configs.xtts_config import XttsConfig from TTS.tts.models.xtts import Xtts from TTS.utils.generic_utils import get_user_data_dir print("application starting") HF_TOKEN = os.environ.get("HF_TOKEN") from huggingface_hub import HfApi # will use api to restart space on a unrecoverable error api = HfApi(token=HF_TOKEN) repo_id = "JacobLinCool/xtts-v2" model = None supported_languages = None def load_model(): global model global supported_languages print("loading model") model_name = "tts_models/multilingual/multi-dataset/xtts_v2" model_path = os.path.join(get_user_data_dir("tts"), model_name.replace("/", "--")) config = XttsConfig() config.load_json(os.path.join(model_path, "config.json")) model = Xtts.init_from_config(config) model.load_checkpoint( config, checkpoint_path=os.path.join(model_path, "model.pth"), vocab_path=os.path.join(model_path, "vocab.json"), eval=True, use_deepspeed=False, ) if torch.cuda.is_available(): model.cuda() else: model.cpu() supported_languages = config.languages print("Model loaded") # This is for debugging purposes only DEVICE_ASSERT_DETECTED = 0 DEVICE_ASSERT_PROMPT = None DEVICE_ASSERT_LANG = None def predict( prompt, language, audio_file_pth, voice_cleanup, no_lang_auto_detect, agree, ): if model is None: load_model() if agree == True: if language not in supported_languages: gr.Warning( f"Language you put {language} in is not in is not in our Supported Languages, please choose from dropdown" ) return ( None, None, None, None, ) language_predicted = langid.classify(prompt)[ 0 ].strip() # strip need as there is space at end! # tts expects chinese as zh-cn if language_predicted == "zh": # we use zh-cn language_predicted = "zh-cn" print(f"Detected language:{language_predicted}, Chosen language:{language}") # After text character length 15 trigger language detection if len(prompt) > 15: # allow any language for short text as some may be common # If user unchecks language autodetection it will not trigger # You may remove this completely for own use if language_predicted != language and not no_lang_auto_detect: # Please duplicate and remove this check if you really want this # Or auto-detector fails to identify language (which it can on pretty short text or mixed text) gr.Warning( f"It looks like your text isn’t the language you chose , if you’re sure the text is the same language you chose, please check disable language auto-detection checkbox" ) return ( None, None, None, None, ) speaker_wav = audio_file_pth # Filtering for microphone input, as it has BG noise, maybe silence in beginning and end # This is fast filtering not perfect # Apply all on demand lowpassfilter = denoise = trim = loudness = True if lowpassfilter: lowpass_highpass = "lowpass=8000,highpass=75," else: lowpass_highpass = "" if trim: # better to remove silence in beginning and end for microphone trim_silence = "areverse,silenceremove=start_periods=1:start_silence=0:start_threshold=0.02,areverse,silenceremove=start_periods=1:start_silence=0:start_threshold=0.02," else: trim_silence = "" if voice_cleanup: try: out_filename = ( speaker_wav + str(uuid.uuid4()) + ".wav" ) # ffmpeg to know output format # we will use newer ffmpeg as that has afftn denoise filter shell_command = f"./ffmpeg -y -i {speaker_wav} -af {lowpass_highpass}{trim_silence} {out_filename}".split( " " ) command_result = subprocess.run( [item for item in shell_command], capture_output=False, text=True, check=True, ) speaker_wav = out_filename print("Filtered microphone input") except subprocess.CalledProcessError: # There was an error - command exited with non-zero code print("Error: failed filtering, use original microphone input") else: speaker_wav = speaker_wav if len(prompt) < 2: gr.Warning("Please give a longer prompt text") return ( None, None, None, None, ) if len(prompt) > 200: gr.Warning( "Text length limited to 200 characters for this demo, please try shorter text. You can clone this space and edit code for your own usage" ) return ( None, None, None, None, ) global DEVICE_ASSERT_DETECTED if DEVICE_ASSERT_DETECTED: global DEVICE_ASSERT_PROMPT global DEVICE_ASSERT_LANG # It will likely never come here as we restart space on first unrecoverable error now print( f"Unrecoverable exception caused by language:{DEVICE_ASSERT_LANG} prompt:{DEVICE_ASSERT_PROMPT}" ) # HF Space specific.. This error is unrecoverable need to restart space space = api.get_space_runtime(repo_id=repo_id) if space.stage != "BUILDING": api.restart_space(repo_id=repo_id) else: print("TRIED TO RESTART but space is building") try: metrics_text = "" t_latent = time.time() # note diffusion_conditioning not used on hifigan (default mode), it will be empty but need to pass it to model.inference try: ( gpt_cond_latent, speaker_embedding, ) = model.get_conditioning_latents( audio_path=speaker_wav, gpt_cond_len=30, max_ref_length=60 ) except Exception as e: print("Speaker encoding error", str(e)) gr.Warning( "It appears something wrong with reference, did you unmute your microphone?" ) return ( None, None, None, None, ) latent_calculation_time = time.time() - t_latent # metrics_text=f"Embedding calculation time: {latent_calculation_time:.2f} seconds\n" # temporary comma fix prompt = re.sub("([^\x00-\x7F]|\w)(\.|\。|\?)", r"\1 \2\2", prompt) wav_chunks = [] ## Direct mode """ print("I: Generating new audio...") t0 = time.time() out = model.inference( prompt, language, gpt_cond_latent, speaker_embedding ) inference_time = time.time() - t0 print(f"I: Time to generate audio: {round(inference_time*1000)} milliseconds") metrics_text+=f"Time to generate audio: {round(inference_time*1000)} milliseconds\n" real_time_factor= (time.time() - t0) / out['wav'].shape[-1] * 24000 print(f"Real-time factor (RTF): {real_time_factor}") metrics_text+=f"Real-time factor (RTF): {real_time_factor:.2f}\n" torchaudio.save("output.wav", torch.tensor(out["wav"]).unsqueeze(0), 24000) """ print("I: Generating new audio in streaming mode...") t0 = time.time() chunks = model.inference_stream( prompt, language, gpt_cond_latent, speaker_embedding, repetition_penalty=7.0, temperature=0.85, ) first_chunk = True for i, chunk in enumerate(chunks): if first_chunk: first_chunk_time = time.time() - t0 metrics_text += f"Latency to first audio chunk: {round(first_chunk_time*1000)} milliseconds\n" first_chunk = False wav_chunks.append(chunk) print(f"Received chunk {i} of audio length {chunk.shape[-1]}") inference_time = time.time() - t0 print( f"I: Time to generate audio: {round(inference_time*1000)} milliseconds" ) # metrics_text += ( # f"Time to generate audio: {round(inference_time*1000)} milliseconds\n" # ) wav = torch.cat(wav_chunks, dim=0) print(wav.shape) real_time_factor = (time.time() - t0) / wav.shape[0] * 24000 print(f"Real-time factor (RTF): {real_time_factor}") metrics_text += f"Real-time factor (RTF): {real_time_factor:.2f}\n" torchaudio.save("output.wav", wav.squeeze().unsqueeze(0).cpu(), 24000) except RuntimeError as e: if "device-side assert" in str(e): # cannot do anything on cuda device side error, need tor estart print( f"Exit due to: Unrecoverable exception caused by language:{language} prompt:{prompt}", flush=True, ) gr.Warning("Unhandled Exception encounter, please retry in a minute") print("Cuda device-assert Runtime encountered need restart") if not DEVICE_ASSERT_DETECTED: DEVICE_ASSERT_DETECTED = 1 DEVICE_ASSERT_PROMPT = prompt DEVICE_ASSERT_LANG = language # just before restarting save what caused the issue so we can handle it in future # Uploading Error data only happens for unrecovarable error error_time = datetime.datetime.now().strftime("%d-%m-%Y-%H:%M:%S") error_data = [ error_time, prompt, language, audio_file_pth, voice_cleanup, no_lang_auto_detect, agree, ] error_data = [str(e) if type(e) != str else e for e in error_data] print(error_data) print(speaker_wav) write_io = StringIO() csv.writer(write_io).writerows([error_data]) csv_upload = write_io.getvalue().encode() filename = error_time + "_" + str(uuid.uuid4()) + ".csv" print("Writing error csv") error_api = HfApi() error_api.upload_file( path_or_fileobj=csv_upload, path_in_repo=filename, repo_id="coqui/xtts-flagged-dataset", repo_type="dataset", ) # speaker_wav print("Writing error reference audio") speaker_filename = ( error_time + "_reference_" + str(uuid.uuid4()) + ".wav" ) error_api = HfApi() error_api.upload_file( path_or_fileobj=speaker_wav, path_in_repo=speaker_filename, repo_id="coqui/xtts-flagged-dataset", repo_type="dataset", ) # HF Space specific.. This error is unrecoverable need to restart space space = api.get_space_runtime(repo_id=repo_id) if space.stage != "BUILDING": api.restart_space(repo_id=repo_id) else: print("TRIED TO RESTART but space is building") else: if "Failed to decode" in str(e): print("Speaker encoding error", str(e)) gr.Warning( "It appears something wrong with reference, did you unmute your microphone?" ) else: print("RuntimeError: non device-side assert error:", str(e)) gr.Warning("Something unexpected happened please retry again.") return ( None, None, None, None, ) return ( gr.make_waveform( audio="output.wav", ), "output.wav", metrics_text, speaker_wav, ) else: gr.Warning("Please accept the Terms & Condition!") return ( None, None, None, None, ) title = "Coqui🐸 XTTS" description = """
XTTS is a text-to-speech model that lets you clone voices into different languages.
This is the same model that powers our creator application Coqui Studio as well as the Coqui API. In production we apply modifications to make low-latency streaming possible.
There are 16 languages.

Arabic: ar, Brazilian Portuguese: pt , Chinese: zh-cn, Czech: cs, Dutch: nl, English: en, French: fr, German: de, Italian: it, Polish: pl, Russian: ru, Spanish: es, Turkish: tr, Japanese: ja, Korean: ko, Hungarian: hu


Leave a star 🌟 on the Github 🐸TTS, where our open-source inference and training code lives.
""" links = """ | | | | ------------------------------- | --------------------------------------- | | 🐸💬 **CoquiTTS** | | | 💼 **Documentation** | [ReadTheDocs](https://tts.readthedocs.io/en/latest/) | 👩‍💻 **Questions** | [GitHub Discussions](https://github.com/coqui-ai/TTS/discussions) | | 🗯 **Community** | [![Dicord](https://img.shields.io/discord/1037326658807533628?color=%239B59B6&label=chat%20on%20discord)](https://discord.gg/5eXr5seRrv) | """ article = """

By using this demo you agree to the terms of the Coqui Public Model License at https://coqui.ai/cpml

We collect data only for error cases for improvement.

""" with gr.Blocks(analytics_enabled=False) as demo: with gr.Row(): with gr.Column(): gr.Markdown( """ ## """ ) with gr.Column(): # placeholder to align the image pass with gr.Row(): with gr.Column(): gr.Markdown(description) with gr.Column(): gr.Markdown(links) with gr.Row(): with gr.Column(): input_text_gr = gr.Textbox( label="Text Prompt", info="One or two sentences at a time is better. Up to 200 text characters.", value="Hi there, I'm your new voice clone. Try your best to upload quality audio", ) language_gr = gr.Dropdown( label="Language", info="Select an output language for the synthesised speech", choices=[ "en", "es", "fr", "de", "it", "pt", "pl", "tr", "ru", "nl", "cs", "ar", "zh-cn", "ja", "ko", "hu", ], value="en", ) ref_gr = gr.Audio( label="Reference Audio", info="Click on the ✎ button to upload your own target speaker audio", type="filepath", value="examples/female.wav", ) clean_ref_gr = gr.Checkbox( label="Cleanup Reference Voice", value=False, info="This check can improve output if your microphone or reference voice is noisy", ) auto_det_lang_gr = gr.Checkbox( label="Do not use language auto-detect", value=False, info="Check to disable language auto-detection", ) tos_gr = gr.Checkbox( label="Agree", value=False, info="I agree to the terms of the Coqui Public Model License at https://coqui.ai/cpml", ) tts_button = gr.Button("Send", elem_id="send-btn", visible=True) with gr.Column(): video_gr = gr.Video(label="Waveform Visual") audio_gr = gr.Audio(label="Synthesised Audio", autoplay=True) out_text_gr = gr.Text(label="Metrics") ref_audio_gr = gr.Audio(label="Reference Audio Used") tts_button.click( predict, [input_text_gr, language_gr, ref_gr, clean_ref_gr, auto_det_lang_gr, tos_gr], outputs=[video_gr, audio_gr, out_text_gr, ref_audio_gr], ) print("Starting server") demo.queue().launch(debug=True, show_api=True)