import io import os import math from queue import Queue from threading import Thread from typing import Optional import numpy as np import spaces import gradio as gr import torch from parler_tts import ParlerTTSForConditionalGeneration from pydub import AudioSegment from transformers import AutoTokenizer, AutoFeatureExtractor, set_seed device = "cuda" if torch.cuda.is_available() else "mps" if torch.backends.mps.is_available() else "cpu" torch_dtype = torch.bfloat16 if device != "cpu" else torch.float32 repo_id = "ai4bharat/indic-parler-tts-pretrained" jenny_repo_id = "ai4bharat/indic-parler-tts" model = ParlerTTSForConditionalGeneration.from_pretrained( repo_id, attn_implementation="eager", torch_dtype=torch_dtype, low_cpu_mem_usage=True ).to(device) jenny_model = ParlerTTSForConditionalGeneration.from_pretrained( jenny_repo_id, attn_implementation="eager", torch_dtype=torch_dtype, low_cpu_mem_usage=True ).to(device) tokenizer = AutoTokenizer.from_pretrained(repo_id) description_tokenizer = AutoTokenizer.from_pretrained("google/flan-t5-large") feature_extractor = AutoFeatureExtractor.from_pretrained(repo_id) SAMPLE_RATE = feature_extractor.sampling_rate SEED = 42 default_text = "Please surprise me and speak in whatever voice you enjoy." examples = [ [ "मुले बागेत खेळत आहेत आणि पक्षी किलबिलाट करत आहेत.", "Sunita speaks slowly in a calm, moderate-pitched voice, delivering the news with a neutral tone. The recording is very high quality with no background noise.", 3.0 ], [ "ಉದ್ಯಾನದಲ್ಲಿ ಮಕ್ಕಳ ಆಟವಾಡುತ್ತಿದ್ದಾರೆ ಮತ್ತು ಪಕ್ಷಿಗಳು ಚಿಲಿಪಿಲಿ ಮಾಡುತ್ತಿವೆ.", "Suresh speaks slowly in a low-pitched, calm voice, with a neutral tone, perfect for narration. The recording is very high quality with no background noise.", 3.0 ], [ "বাচ্চারা বাগানে খেলছে আর পাখি কিচিরমিচির করছে।", "Aditi speaks at a moderate pace and pitch, with a clear, neutral tone and no emotional emphasis. The recording is very high quality with no background noise.", 3.0 ], [ "పిల్లలు తోటలో ఆడుకుంటున్నారు, పక్షుల కిలకిలరావాలు.", "Prakash speaks slowly in a low-pitched, calm voice, with a neutral tone, perfect for narration. The recording is very high quality with no background noise.", 3.0 ], [ "పిల్లలు తోటలో ఆడుకుంటున్నారు, పక్షుల కిలకిలరావాలు.", "Prakash speaks slowly in a low-pitched, calm voice, with a neutral tone, perfect for narration. The recording is very high quality with no background noise.", 3.0 ], [ "This is the best time of my life, Bartley,' she said happily", "A male speaker with a low-pitched voice speaks with a British accent at a fast pace in a small, confined space with very clear audio and an animated tone.", 3.0 ], [ "Montrose also, after having experienced still more variety of good and bad fortune, threw down his arms, and retired out of the kingdom.", "A female speaker with a slightly low-pitched, quite monotone voice speaks with an American accent at a slightly faster-than-average pace in a confined space with very clear audio.", 3.0 ], [ "बगीचे में बच्चे खेल रहे हैं और पक्षी चहचहा रहे हैं।", "Rohit speaks with a slightly high-pitched voice delivering his words at a slightly slow pace in a small, confined space with a touch of background noise and a quite monotone tone.", 3.0 ], [ "കുട്ടികൾ പൂന്തോട്ടത്തിൽ കളിക്കുന്നു, പക്ഷികൾ ചിലയ്ക്കുന്നു.", "Anjali speaks with a low-pitched voice delivering her words at a fast pace and an animated tone, in a very spacious environment, accompanied by noticeable background noise.", 3.0 ], [ "குழந்தைகள் தோட்டத்தில் விளையாடுகிறார்கள், பறவைகள் கிண்டல் செய்கின்றன.", "Jaya speaks with a slightly low-pitched, quite monotone voice at a slightly faster-than-average pace in a confined space with very clear audio.", 3.0 ] ] jenny_examples = [ [ "मुले बागेत खेळत आहेत आणि पक्षी किलबिलाट करत आहेत.", "Sunita speaks slowly in a calm, moderate-pitched voice, delivering the news with a neutral tone. The recording is very high quality with no background noise.", 3.0 ], [ "ಉದ್ಯಾನದಲ್ಲಿ ಮಕ್ಕಳ ಆಟವಾಡುತ್ತಿದ್ದಾರೆ ಮತ್ತು ಪಕ್ಷಿಗಳು ಚಿಲಿಪಿಲಿ ಮಾಡುತ್ತಿವೆ.", "Suresh speaks slowly in a low-pitched, calm voice, with a neutral tone, perfect for narration. The recording is very high quality with no background noise.", 3.0 ], [ "বাচ্চারা বাগানে খেলছে আর পাখি কিচিরমিচির করছে।", "Aditi speaks at a moderate pace and pitch, with a clear, neutral tone and no emotional emphasis. The recording is very high quality with no background noise.", 3.0 ], [ "పిల్లలు తోటలో ఆడుకుంటున్నారు, పక్షుల కిలకిలరావాలు.", "Prakash speaks slowly in a low-pitched, calm voice, with a neutral tone, perfect for narration. The recording is very high quality with no background noise.", 3.0 ], [ "పిల్లలు తోటలో ఆడుకుంటున్నారు, పక్షుల కిలకిలరావాలు.", "Prakash speaks slowly in a low-pitched, calm voice, with a neutral tone, perfect for narration. The recording is very high quality with no background noise.", 3.0 ], [ "This is the best time of my life, Bartley,' she said happily", "A male speaker with a low-pitched voice speaks with a British accent at a fast pace in a small, confined space with very clear audio and an animated tone.", 3.0 ], [ "Montrose also, after having experienced still more variety of good and bad fortune, threw down his arms, and retired out of the kingdom.", "A female speaker with a slightly low-pitched, quite monotone voice speaks with an American accent at a slightly faster-than-average pace in a confined space with very clear audio.", 3.0 ], [ "बगीचे में बच्चे खेल रहे हैं और पक्षी चहचहा रहे हैं।", "Rohit speaks with a slightly high-pitched voice delivering his words at a slightly slow pace in a small, confined space with a touch of background noise and a quite monotone tone.", 3.0 ], [ "കുട്ടികൾ പൂന്തോട്ടത്തിൽ കളിക്കുന്നു, പക്ഷികൾ ചിലയ്ക്കുന്നു.", "Anjali speaks with a low-pitched voice delivering her words at a fast pace and an animated tone, in a very spacious environment, accompanied by noticeable background noise.", 3.0 ], [ "குழந்தைகள் தோட்டத்தில் விளையாடுகிறார்கள், பறவைகள் கிண்டல் செய்கின்றன.", "Jaya speaks with a slightly low-pitched, quite monotone voice at a slightly faster-than-average pace in a confined space with very clear audio.", 3.0 ] ] def numpy_to_mp3(audio_array, sampling_rate): # Normalize audio_array if it's floating-point if np.issubdtype(audio_array.dtype, np.floating): max_val = np.max(np.abs(audio_array)) audio_array = (audio_array / max_val) * 32767 # Normalize to 16-bit range audio_array = audio_array.astype(np.int16) # Create an audio segment from the numpy array audio_segment = AudioSegment( audio_array.tobytes(), frame_rate=sampling_rate, sample_width=audio_array.dtype.itemsize, channels=1 ) # Export the audio segment to MP3 bytes - use a high bitrate to maximise quality mp3_io = io.BytesIO() audio_segment.export(mp3_io, format="mp3", bitrate="320k") # Get the MP3 bytes mp3_bytes = mp3_io.getvalue() mp3_io.close() return mp3_bytes sampling_rate = model.audio_encoder.config.sampling_rate frame_rate = model.audio_encoder.config.frame_rate # @spaces.GPU # def generate_base(text, description, play_steps_in_s=2.0): # play_steps = int(frame_rate * play_steps_in_s) # streamer = ParlerTTSStreamer(model, device=device, play_steps=play_steps) # inputs = description_tokenizer(description, return_tensors="pt").to(device) # prompt = tokenizer(text, return_tensors="pt").to(device) # generation_kwargs = dict( # input_ids=inputs.input_ids, # prompt_input_ids=prompt.input_ids, # streamer=streamer, # do_sample=True, # temperature=1.0, # min_new_tokens=10, # ) # set_seed(SEED) # thread = Thread(target=model.generate, kwargs=generation_kwargs) # thread.start() # for new_audio in streamer: # print(f"Sample of length: {round(new_audio.shape[0] / sampling_rate, 2)} seconds") # yield numpy_to_mp3(new_audio, sampling_rate=sampling_rate) @spaces.GPU def generate_base(text, description, play_steps_in_s=2.0): # Initialize variables play_steps = int(frame_rate * play_steps_in_s) chunk_size = 15 # Process 10 words at a time # Tokenize the full text and description inputs = description_tokenizer(description, return_tensors="pt").to(device) # Split text into chunks of approximately 10 words words = text.split() chunks = [' '.join(words[i:i + chunk_size]) for i in range(0, len(words), chunk_size)] all_audio = [] # Process each chunk for chunk in chunks: # Tokenize the chunk prompt = tokenizer(chunk, return_tensors="pt").to(device) # Generate audio for the chunk generation = model.generate( input_ids=inputs.input_ids, attention_mask=inputs.attention_mask, prompt_input_ids=prompt.input_ids, prompt_attention_mask=prompt.attention_mask, do_sample=True, # temperature=1.0, # min_new_tokens=10, return_dict_in_generate=True ) # Extract audio from generation if hasattr(generation, 'sequences') and hasattr(generation, 'audios_length'): audio = generation.sequences[0, :generation.audios_length[0]] audio_np = audio.to(torch.float32).cpu().numpy().squeeze() if len(audio_np.shape) > 1: audio_np = audio_np.flatten() all_audio.append(audio_np) # Combine all audio chunks combined_audio = np.concatenate(all_audio) # Convert to expected format and yield print(f"Sample of length: {round(combined_audio.shape[0] / sampling_rate, 2)} seconds") yield numpy_to_mp3(combined_audio, sampling_rate=sampling_rate) # @spaces.GPU # def generate_jenny(text, description, play_steps_in_s=2.0): # play_steps = int(frame_rate * play_steps_in_s) # streamer = ParlerTTSStreamer(jenny_model, device=device, play_steps=play_steps) # inputs = description_tokenizer(description, return_tensors="pt").to(device) # prompt = tokenizer(text, return_tensors="pt").to(device) # generation_kwargs = dict( # input_ids=inputs.input_ids, # prompt_input_ids=prompt.input_ids, # streamer=streamer, # do_sample=True, # temperature=1.0, # min_new_tokens=10, # ) # set_seed(SEED) # thread = Thread(target=jenny_model.generate, kwargs=generation_kwargs) # thread.start() # for new_audio in streamer: # print(f"Sample of length: {round(new_audio.shape[0] / sampling_rate, 2)} seconds") # yield sampling_rate, new_audio @spaces.GPU def generate_jenny(text, description, play_steps_in_s=2.0): # Initialize variables play_steps = int(frame_rate * play_steps_in_s) chunk_size = 15 # Process 10 words at a time # Tokenize the full text and description inputs = description_tokenizer(description, return_tensors="pt").to(device) # Split text into chunks of approximately 10 words words = text.split() chunks = [' '.join(words[i:i + chunk_size]) for i in range(0, len(words), chunk_size)] all_audio = [] # Process each chunk for chunk in chunks: # Tokenize the chunk prompt = tokenizer(chunk, return_tensors="pt").to(device) # Generate audio for the chunk generation = jenny_model.generate( input_ids=inputs.input_ids, attention_mask=inputs.attention_mask, prompt_input_ids=prompt.input_ids, prompt_attention_mask=prompt.attention_mask, do_sample=True, # temperature=1.0, # min_new_tokens=10, return_dict_in_generate=True ) # Extract audio from generation if hasattr(generation, 'sequences') and hasattr(generation, 'audios_length'): audio = generation.sequences[0, :generation.audios_length[0]] audio_np = audio.to(torch.float32).cpu().numpy().squeeze() if len(audio_np.shape) > 1: audio_np = audio_np.flatten() all_audio.append(audio_np) # Combine all audio chunks combined_audio = np.concatenate(all_audio) # Convert to expected format and yield print(f"Sample of length: {round(combined_audio.shape[0] / sampling_rate, 2)} seconds") yield numpy_to_mp3(combined_audio, sampling_rate=sampling_rate) 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( """

Parler-TTS 🗣️

""" ) gr.HTML( f"""

IndicParlerTTS is a training and inference library for high-quality text-to-speech (TTS) models. This demonstration highlights the flexibility of the IndicParlerTTS model, which generates natural, expressive speech for over 22 Indian languages, using a simple text prompt to control features like speaker style, tone, pitch, pace, and more.

Tips for effective usage:

""" ) with gr.Tab("Finetuned"): with gr.Row(): with gr.Column(): input_text = gr.Textbox(label="Input Text", lines=2, value=jenny_examples[0][0], elem_id="input_text") description = gr.Textbox(label="Description", lines=2, value=jenny_examples[0][1], elem_id="input_description") play_seconds = gr.Slider(3.0, 7.0, value=jenny_examples[0][2], step=2, label="Streaming interval in seconds", info="Lower = shorter chunks, lower latency, more codec steps") run_button = gr.Button("Generate Audio", variant="primary") with gr.Column(): audio_out = gr.Audio(label="Parler-TTS generation", format="mp3", elem_id="audio_out", streaming=True, autoplay=True) inputs = [input_text, description, play_seconds] outputs = [audio_out] gr.Examples(examples=jenny_examples, fn=generate_jenny, inputs=inputs, outputs=outputs, cache_examples=False) run_button.click(fn=generate_jenny, inputs=inputs, outputs=outputs, queue=True) with gr.Tab("Pretrained"): 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="", elem_id="input_description") play_seconds = gr.Slider(3.0, 7.0, value=3.0, step=2, label="Streaming interval in seconds", info="Lower = shorter chunks, lower latency, more codec steps") run_button = gr.Button("Generate Audio", variant="primary") with gr.Column(): audio_out = gr.Audio(label="Parler-TTS generation", format="mp3", elem_id="audio_out", streaming=True, autoplay=True) inputs = [input_text, description, play_seconds] outputs = [audio_out] gr.Examples(examples=examples, fn=generate_base, inputs=inputs, outputs=outputs, cache_examples=False) run_button.click(fn=generate_base, inputs=inputs, outputs=outputs, queue=True) gr.HTML( """ If you'd like to learn more about how the model was trained or explore fine-tuning it yourself, visit the Parler-TTS repository on GitHub. The Parler-TTS codebase and associated checkpoints are licensed under the Apache 2.0 license.

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