import gradio as gr import random import os import shutil import pandas as pd import sqlite3 from datasets import load_dataset import threading import time import uuid from pathlib import Path from huggingface_hub import CommitScheduler, delete_file, hf_hub_download from gradio_client import Client from detoxify import Detoxify toxicity = Detoxify('original') with open('harvard_sentences.txt') as f: sents = f.read().strip().splitlines() #################################### # Constants #################################### AVAILABLE_MODELS = { 'XTTS': 'xttsv2', 'WhisperSpeech': 'whisperspeech', 'ElevenLabs': 'eleven', 'OpenVoice': 'openvoice', 'Pheme': 'pheme', 'MetaVoice': 'metavoice', 'OpenAI TTS': 'openai', } SPACE_ID = os.getenv('HF_ID') MAX_SAMPLE_TXT_LENGTH = 150 MIN_SAMPLE_TXT_LENGTH = 10 DB_DATASET_ID = os.getenv('DATASET_ID') DB_NAME = "database.db" # If /data available => means local storage is enabled => let's use it! DB_PATH = f"/data/{DB_NAME}" if os.path.isdir("/data") else DB_NAME print(f"Using {DB_PATH}") # AUDIO_DATASET_ID = "ttseval/tts-arena-new" CITATION_TEXT = """@misc{tts-arena, title = {Text to Speech Arena}, author = {mrfakename and Srivastav, Vaibhav and Pouget, Lucain and Fourrier, Clémentine}, year = 2024, publisher = {Hugging Face}, howpublished = "\\url{https://huggingface.co/spaces/ttseval/TTS-Arena}" }""" #################################### # Functions #################################### def create_db_if_missing(): conn = get_db() cursor = conn.cursor() cursor.execute(''' CREATE TABLE IF NOT EXISTS model ( name TEXT UNIQUE, upvote INTEGER, downvote INTEGER ); ''') cursor.execute(''' CREATE TABLE IF NOT EXISTS vote ( id INTEGER PRIMARY KEY AUTOINCREMENT, username TEXT, model TEXT, vote INTEGER ); ''') cursor.execute(''' CREATE TABLE IF NOT EXISTS spokentext ( id INTEGER PRIMARY KEY AUTOINCREMENT, spokentext TEXT ); ''') def get_db(): return sqlite3.connect(DB_PATH) #################################### # Space initialization #################################### # Download existing DB if not os.path.isfile(DB_PATH): print("Downloading DB...") try: cache_path = hf_hub_download(repo_id=DB_DATASET_ID, repo_type='dataset', filename=DB_NAME) shutil.copyfile(cache_path, DB_PATH) print("Downloaded DB") except Exception as e: print("Error while downloading DB:", e) # Create DB table (if doesn't exist) create_db_if_missing() # Sync local DB with remote repo every 5 minute (only if a change is detected) scheduler = CommitScheduler( repo_id=DB_DATASET_ID, repo_type="dataset", folder_path=Path(DB_PATH).parent, every=5, allow_patterns=DB_NAME, ) # Load audio dataset # audio_dataset = load_dataset(AUDIO_DATASET_ID) #################################### # Router API #################################### router = Client("ttseval/tts-router", hf_token=os.getenv('HF_TOKEN')) #################################### # Gradio app #################################### MUST_BE_LOGGEDIN = "Please login with Hugging Face to participate in the TTS Arena." DESCR = """ # TTS Arena Vote on different speech synthesis models! """.strip() # INSTR = """ # ## Instructions # * Listen to two anonymous models # * Vote on which synthesized audio sounds more natural to you # * If there's a tie, click Skip # **When you're ready to begin, login and begin voting!** The model names will be revealed once you vote. # """.strip() INSTR = """ ## Instructions * Enter text to synthesize. * Listen to the two audio clips. * Vote on which synthesized audio sounds more natural to you. **Did the model hallucinate?** **When you're ready to begin, enter text!** The model names will be revealed once you vote. """.strip() request = '' if SPACE_ID: request = f""" ### Request Model Please fill out [this form](https://huggingface.co/spaces/{SPACE_ID}/discussions/new?title=%5BModel+Request%5D+&description=%23%23%20Model%20Request%0A%0A%2A%2AModel%20website%2Fpaper%20%28if%20applicable%29%2A%2A%3A%0A%2A%2AModel%20available%20on%2A%2A%3A%20%28coqui%7CHF%20pipeline%7Ccustom%20code%29%0A%2A%2AWhy%20do%20you%20want%20this%20model%20added%3F%2A%2A%0A%2A%2AComments%3A%2A%2A) to request a model. """ ABOUT = f""" ## About The TTS Arena is a project created to evaluate leading speech synthesis models. It is inspired by the [Chatbot Arena](https://chat.lmsys.org/) by LMSYS. ### How it Works First, vote on two samples of text-to-speech models. The models that synthesized the samples are not revealed to mitigate bias. As you vote, the leaderboard will be updated based on votes. We calculate a score for each model using a method similar to the [Elo system](https://en.wikipedia.org/wiki/Elo_rating_system). ### Motivation Recently, many new open-access speech synthesis models have been made available to the community. However, there is no standardized evaluation or benchmark to measure the quality and naturalness of these models. The TTS Arena is an attempt to benchmark these models and find the highest-quality models available to the community. {request} ### Privacy Statement We may store text you enter and generated audio. We store a unique ID for each session. ### License Please assume all generated audio clips are not licensed to be redistributed and may only be used for personal, non-commercial use. """.strip() LDESC = """ ## Leaderboard A list of the models, based on how highly they are ranked! ### **Important**: To keep a fair impression of results, the leaderboard will be **hidden** by default, until a large number of human ratings have been recorded. Tick the `Reveal Preliminary Results` checkbox below if you wish to see the raw data. """.strip() # def reload_audio_dataset(): # global audio_dataset # audio_dataset = load_dataset(AUDIO_DATASET_ID) # return 'Reload Audio Dataset' def del_db(txt): if not txt.lower() == 'delete db': raise gr.Error('You did not enter "delete db"') # Delete local + remote os.remove(DB_PATH) delete_file(path_in_repo=DB_NAME, repo_id=DB_DATASET_ID, repo_type='dataset') # Recreate create_db_if_missing() return 'Delete DB' theme = gr.themes.Base( font=[gr.themes.GoogleFont('Libre Franklin'), gr.themes.GoogleFont('Public Sans'), 'system-ui', 'sans-serif'], ) model_names = { 'styletts2': 'StyleTTS 2', 'tacotron': 'Tacotron', 'tacotronph': 'Tacotron Phoneme', 'tacotrondca': 'Tacotron DCA', 'speedyspeech': 'Speedy Speech', 'overflow': 'Overflow TTS', 'vits': 'VITS', 'vitsneon': 'VITS Neon', 'neuralhmm': 'Neural HMM', 'glow': 'Glow TTS', 'fastpitch': 'FastPitch', 'jenny': 'Jenny', 'tortoise': 'Tortoise TTS', 'xtts2': 'Coqui XTTSv2', 'xtts': 'Coqui XTTS', 'openvoice': 'MyShell OpenVoice', 'elevenlabs': 'ElevenLabs', 'openai': 'OpenAI', 'hierspeech': 'HierSpeech++', 'pheme': 'PolyAI Pheme', 'speecht5': 'SpeechT5', 'metavoice': 'MetaVoice-1B', } model_licenses = { 'styletts2': 'MIT', 'tacotron': 'BSD-3', 'tacotronph': 'BSD-3', 'tacotrondca': 'BSD-3', 'speedyspeech': 'BSD-3', 'overflow': 'MIT', 'vits': 'MIT', 'openvoice': 'MIT', 'vitsneon': 'BSD-3', 'neuralhmm': 'MIT', 'glow': 'MIT', 'fastpitch': 'Apache 2.0', 'jenny': 'Jenny License', 'tortoise': 'Apache 2.0', 'xtts2': 'CPML (NC)', 'xtts': 'CPML (NC)', 'elevenlabs': 'Proprietary', 'openai': 'Proprietary', 'hierspeech': 'MIT', 'pheme': 'CC-BY', 'speecht5': 'MIT', 'metavoice': 'Apache 2.0', 'elevenlabs': 'Proprietary', } model_links = { 'styletts2': 'https://github.com/yl4579/StyleTTS2', 'tacotron': 'https://github.com/NVIDIA/tacotron2', 'speedyspeech': 'https://github.com/janvainer/speedyspeech', 'overflow': 'https://github.com/shivammehta25/OverFlow', 'vits': 'https://github.com/jaywalnut310/vits', 'openvoice': 'https://github.com/myshell-ai/OpenVoice', 'neuralhmm': 'https://github.com/ketranm/neuralHMM', 'glow': 'https://github.com/jaywalnut310/glow-tts', 'fastpitch': 'https://fastpitch.github.io/', 'tortoise': 'https://github.com/neonbjb/tortoise-tts', 'xtts2': 'https://huggingface.co/coqui/XTTS-v2', 'xtts': 'https://huggingface.co/coqui/XTTS-v1', 'elevenlabs': 'https://elevenlabs.io/', 'openai': 'https://help.openai.com/en/articles/8555505-tts-api', 'hierspeech': 'https://github.com/sh-lee-prml/HierSpeechpp', 'pheme': 'https://github.com/PolyAI-LDN/pheme', 'speecht5': 'https://github.com/microsoft/SpeechT5', 'metavoice': 'https://github.com/metavoiceio/metavoice-src', } # def get_random_split(existing_split=None): # choice = random.choice(list(audio_dataset.keys())) # if existing_split and choice == existing_split: # return get_random_split(choice) # else: # return choice # def get_random_splits(): # choice1 = get_random_split() # choice2 = get_random_split(choice1) # return (choice1, choice2) def model_license(name): if name in model_licenses.keys(): return model_licenses[name] return 'Unknown' def get_leaderboard(reveal_prelim: bool): conn = get_db() cursor = conn.cursor() sql = 'SELECT name, upvote, downvote FROM model' if not reveal_prelim: sql += ' WHERE EXISTS (SELECT 1 FROM model WHERE (upvote + downvote) > 750)' cursor.execute(sql) data = cursor.fetchall() df = pd.DataFrame(data, columns=['name', 'upvote', 'downvote']) df['license'] = df['name'].map(model_license).fillna("Unknown") df['name'] = df['name'].replace(model_names) df['votes'] = df['upvote'] + df['downvote'] # df['score'] = round((df['upvote'] / df['votes']) * 100, 2) # Percentage score ## ELO SCORE df['score'] = 1200 for i in range(len(df)): for j in range(len(df)): if i != j: expected_a = 1 / (1 + 10 ** ((df['score'][j] - df['score'][i]) / 400)) expected_b = 1 / (1 + 10 ** ((df['score'][i] - df['score'][j]) / 400)) actual_a = df['upvote'][i] / df['votes'][i] actual_b = df['upvote'][j] / df['votes'][j] df.at[i, 'score'] += 32 * (actual_a - expected_a) df.at[j, 'score'] += 32 * (actual_b - expected_b) df['score'] = round(df['score']) ## ELO SCORE df = df.sort_values(by='score', ascending=False) df['order'] = ['#' + str(i + 1) for i in range(len(df))] # df = df[['name', 'score', 'upvote', 'votes']] df = df[['order', 'name', 'score', 'license', 'votes']] return df def mkuuid(uid): if not uid: uid = uuid.uuid4() return uid def upvote_model(model, uname): conn = get_db() cursor = conn.cursor() cursor.execute('UPDATE model SET upvote = upvote + 1 WHERE name = ?', (model,)) if cursor.rowcount == 0: cursor.execute('INSERT OR REPLACE INTO model (name, upvote, downvote) VALUES (?, 1, 0)', (model,)) cursor.execute('INSERT INTO vote (username, model, vote) VALUES (?, ?, ?)', (uname, model, 1,)) with scheduler.lock: conn.commit() cursor.close() def log_text(text): conn = get_db() cursor = conn.cursor() cursor.execute('INSERT INTO spokentext (spokentext) VALUES (?)', (text,)) with scheduler.lock: conn.commit() cursor.close() def downvote_model(model, uname): conn = get_db() cursor = conn.cursor() cursor.execute('UPDATE model SET downvote = downvote + 1 WHERE name = ?', (model,)) if cursor.rowcount == 0: cursor.execute('INSERT OR REPLACE INTO model (name, upvote, downvote) VALUES (?, 0, 1)', (model,)) cursor.execute('INSERT INTO vote (username, model, vote) VALUES (?, ?, ?)', (uname, model, -1,)) with scheduler.lock: conn.commit() cursor.close() def a_is_better(model1, model2, userid): userid = mkuuid(userid) if model1 and model2: upvote_model(model1, str(userid)) downvote_model(model2, str(userid)) return reload(model1, model2, userid, chose_a=True) def b_is_better(model1, model2, userid): userid = mkuuid(userid) if model1 and model2: upvote_model(model2, str(userid)) downvote_model(model1, str(userid)) return reload(model1, model2, userid, chose_b=True) def both_bad(model1, model2, userid): userid = mkuuid(userid) if model1 and model2: downvote_model(model1, str(userid)) downvote_model(model2, str(userid)) return reload(model1, model2, userid) def both_good(model1, model2, userid): userid = mkuuid(userid) if model1 and model2: upvote_model(model1, str(userid)) upvote_model(model2, str(userid)) return reload(model1, model2, userid) def reload(chosenmodel1=None, chosenmodel2=None, userid=None, chose_a=False, chose_b=False): # Select random splits # row = random.choice(list(audio_dataset['train'])) # options = list(random.choice(list(audio_dataset['train'])).keys()) # split1, split2 = random.sample(options, 2) # choice1, choice2 = (row[split1], row[split2]) # if chosenmodel1 in model_names: # chosenmodel1 = model_names[chosenmodel1] # if chosenmodel2 in model_names: # chosenmodel2 = model_names[chosenmodel2] # out = [ # (choice1['sampling_rate'], choice1['array']), # (choice2['sampling_rate'], choice2['array']), # split1, # split2 # ] # if userid: out.append(userid) # if chosenmodel1: out.append(f'This model was {chosenmodel1}') # if chosenmodel2: out.append(f'This model was {chosenmodel2}') # return out # return (f'This model was {chosenmodel1}', f'This model was {chosenmodel2}', gr.update(visible=False), gr.update(visible=False)) # return (gr.update(variant='secondary', value=chosenmodel1, interactive=False), gr.update(variant='secondary', value=chosenmodel2, interactive=False)) out = [ gr.update(interactive=False, visible=False), gr.update(interactive=False, visible=False) ] if chose_a == True: out.append(gr.update(value=f'Your vote: {chosenmodel1}', interactive=False, visible=True)) out.append(gr.update(value=f'{chosenmodel2}', interactive=False, visible=True)) else: out.append(gr.update(value=f'{chosenmodel1}', interactive=False, visible=True)) out.append(gr.update(value=f'Your vote: {chosenmodel2}', interactive=False, visible=True)) out.append(gr.update(visible=True)) return out with gr.Blocks() as leaderboard: gr.Markdown(LDESC) # df = gr.Dataframe(interactive=False, value=get_leaderboard()) df = gr.Dataframe(interactive=False, min_width=0, wrap=True, column_widths=[30, 200, 50, 75, 50]) with gr.Row(): reveal_prelim = gr.Checkbox(label="Reveal Preliminary Results", info="Show all models, including models with very few human ratings.", scale=0) reloadbtn = gr.Button("Refresh") reveal_prelim.input(get_leaderboard, inputs=[reveal_prelim], outputs=[df]) leaderboard.load(get_leaderboard, inputs=[reveal_prelim], outputs=[df]) reloadbtn.click(get_leaderboard, inputs=[reveal_prelim], outputs=[df]) gr.Markdown("DISCLAIMER: The licenses listed may not be accurate or up to date, you are responsible for checking the licenses before using the models. Also note that some models may have additional usage restrictions.") # with gr.Blocks() as vote: # useridstate = gr.State() # gr.Markdown(INSTR) # # gr.LoginButton() # with gr.Row(): # gr.HTML('

Model A

') # gr.HTML('

Model B

') # model1 = gr.Textbox(interactive=False, visible=False, lines=1, max_lines=1) # model2 = gr.Textbox(interactive=False, visible=False, lines=1, max_lines=1) # # with gr.Group(): # # with gr.Row(): # # prevmodel1 = gr.Textbox(interactive=False, show_label=False, container=False, value="Vote to reveal model A") # # prevmodel2 = gr.Textbox(interactive=False, show_label=False, container=False, value="Vote to reveal model B", text_align="right") # # with gr.Row(): # # aud1 = gr.Audio(interactive=False, show_label=False, show_download_button=False, show_share_button=False, waveform_options={'waveform_progress_color': '#3C82F6'}) # # aud2 = gr.Audio(interactive=False, show_label=False, show_download_button=False, show_share_button=False, waveform_options={'waveform_progress_color': '#3C82F6'}) # with gr.Group(): # with gr.Row(): # with gr.Column(): # with gr.Group(): # prevmodel1 = gr.Textbox(interactive=False, show_label=False, container=False, value="Vote to reveal model A", lines=1, max_lines=1) # aud1 = gr.Audio(interactive=False, show_label=False, show_download_button=False, show_share_button=False, waveform_options={'waveform_progress_color': '#3C82F6'}) # with gr.Column(): # with gr.Group(): # prevmodel2 = gr.Textbox(interactive=False, show_label=False, container=False, value="Vote to reveal model B", text_align="right", lines=1, max_lines=1) # aud2 = gr.Audio(interactive=False, show_label=False, show_download_button=False, show_share_button=False, waveform_options={'waveform_progress_color': '#3C82F6'}) # with gr.Row(): # abetter = gr.Button("A is Better", variant='primary', scale=4) # # skipbtn = gr.Button("Skip", scale=1) # bbetter = gr.Button("B is Better", variant='primary', scale=4) # with gr.Row(): # bothbad = gr.Button("Both are Bad", scale=2) # skipbtn = gr.Button("Skip", scale=1) # bothgood = gr.Button("Both are Good", scale=2) # outputs = [aud1, aud2, model1, model2, useridstate, prevmodel1, prevmodel2] # abetter.click(a_is_better, outputs=outputs, inputs=[model1, model2, useridstate]) # bbetter.click(b_is_better, outputs=outputs, inputs=[model1, model2, useridstate]) # skipbtn.click(b_is_better, outputs=outputs, inputs=[model1, model2, useridstate]) # bothbad.click(both_bad, outputs=outputs, inputs=[model1, model2, useridstate]) # bothgood.click(both_good, outputs=outputs, inputs=[model1, model2, useridstate]) # vote.load(reload, outputs=[aud1, aud2, model1, model2]) def synthandreturn(text): text = text.strip() if len(text) > MAX_SAMPLE_TXT_LENGTH: raise gr.Error(f'You exceeded the limit of {MAX_SAMPLE_TXT_LENGTH} characters') if len(text) < MIN_SAMPLE_TXT_LENGTH: raise gr.Error(f'Not enough text') if (toxicity.predict(text)['toxicity'] > 0.5): raise gr.Error('Your text failed the toxicity test!') if not text: raise gr.Error(f'You did not enter any text') # Get two random models mdl1, mdl2 = random.sample(list(AVAILABLE_MODELS.keys()), 2) log_text(text) print("[debug] Using", mdl1, mdl2) return ( text, "Synthesize", gr.update(visible=True), # r2 mdl1, # model1 mdl2, # model2 # 'Vote to reveal model A', # prevmodel1 gr.update(visible=True, value=router.predict( text, AVAILABLE_MODELS[mdl1], api_name="/synthesize" )), # aud1 # 'Vote to reveal model B', # prevmodel2 gr.update(visible=True, value=router.predict( text, AVAILABLE_MODELS[mdl2], api_name="/synthesize" )), # aud2 gr.update(visible=True, interactive=True), gr.update(visible=True, interactive=True), gr.update(visible=False), gr.update(visible=False), gr.update(visible=False), #nxt round btn ) def randomsent(): return random.choice(sents), '🎲' def clear_stuff(): return "", "Synthesize", gr.update(visible=False), '', '', gr.update(visible=False), gr.update(visible=False), gr.update(visible=False), gr.update(visible=False), gr.update(visible=False), gr.update(visible=False), gr.update(visible=False) with gr.Blocks() as vote: useridstate = gr.State() gr.Markdown(INSTR) with gr.Group(): with gr.Row(): text = gr.Textbox(container=False, show_label=False, placeholder="Enter text to synthesize", lines=1, max_lines=1, scale=9999999, min_width=0) randomt = gr.Button('🎲', scale=0, min_width=0) randomt.click(randomsent, outputs=[text, randomt]) btn = gr.Button("Synthesize", variant='primary') model1 = gr.Textbox(interactive=False, lines=1, max_lines=1, visible=False) model2 = gr.Textbox(interactive=False, lines=1, max_lines=1, visible=False) with gr.Row(visible=False) as r2: with gr.Column(): with gr.Group(): aud1 = gr.Audio(interactive=False, show_label=False, show_download_button=False, show_share_button=False, waveform_options={'waveform_progress_color': '#3C82F6'}) abetter = gr.Button("A is better", variant='primary') prevmodel1 = gr.Textbox(interactive=False, show_label=False, container=False, value="Vote to reveal model A", text_align="center", lines=1, max_lines=1, visible=False) with gr.Column(): with gr.Group(): aud2 = gr.Audio(interactive=False, show_label=False, show_download_button=False, show_share_button=False, waveform_options={'waveform_progress_color': '#3C82F6'}) bbetter = gr.Button("B is better", variant='primary') prevmodel2 = gr.Textbox(interactive=False, show_label=False, container=False, value="Vote to reveal model B", text_align="center", lines=1, max_lines=1, visible=False) nxtroundbtn = gr.Button('Next round', visible=False) # outputs = [text, btn, r2, model1, model2, prevmodel1, aud1, prevmodel2, aud2, abetter, bbetter] outputs = [text, btn, r2, model1, model2, aud1, aud2, abetter, bbetter, prevmodel1, prevmodel2, nxtroundbtn] btn.click(synthandreturn, inputs=[text], outputs=outputs) nxtroundbtn.click(clear_stuff, outputs=outputs) # nxt_outputs = [prevmodel1, prevmodel2, abetter, bbetter] nxt_outputs = [abetter, bbetter, prevmodel1, prevmodel2, nxtroundbtn] abetter.click(a_is_better, outputs=nxt_outputs, inputs=[model1, model2, useridstate]) bbetter.click(b_is_better, outputs=nxt_outputs, inputs=[model1, model2, useridstate]) # skipbtn.click(b_is_better, outputs=outputs, inputs=[model1, model2, useridstate]) # bothbad.click(both_bad, outputs=outputs, inputs=[model1, model2, useridstate]) # bothgood.click(both_good, outputs=outputs, inputs=[model1, model2, useridstate]) # vote.load(reload, outputs=[aud1, aud2, model1, model2]) with gr.Blocks() as about: gr.Markdown(ABOUT) with gr.Blocks() as admin: rdb = gr.Button("Reload Audio Dataset") # rdb.click(reload_audio_dataset, outputs=rdb) with gr.Group(): dbtext = gr.Textbox(label="Type \"delete db\" to confirm", placeholder="delete db") ddb = gr.Button("Delete DB") ddb.click(del_db, inputs=dbtext, outputs=ddb) with gr.Blocks(theme=theme, css="footer {visibility: hidden}textbox{resize:none}", title="TTS Leaderboard") as demo: gr.Markdown(DESCR) gr.TabbedInterface([vote, leaderboard, about, admin], ['Vote', 'Leaderboard', 'About', 'Admin (ONLY IN BETA)']) if CITATION_TEXT: with gr.Row(): with gr.Accordion("Citation", open=False): gr.Markdown(f"If you use this data in your publication, please cite us!\n\nCopy the BibTeX citation to cite this source:\n\n```bibtext\n{CITATION_TEXT}\n```\n\nPlease remember that all generated audio clips should be assumed unsuitable for redistribution or commercial use.") demo.queue(api_open=False).launch(show_api=False)