neural-sync 0.1.0

A library to standardize the usage of various machine learning models

Neural_sync Library

Overview

The Neural_sync Library provides a unified interface for working with different machine learning models across various tasks. This library aims to standardize the way models are loaded, parameters are set, and results are generated, enabling a consistent approach regardless of the model type.

Supported Models

Text-to-Speech (TTS)

• FastPitch & HiFi-GAN: Convert text to high-quality speech audio.

Speech-to-Text (STT)

• Distil-Large-V2: Transcribe speech to text.
• Openai/whisper-large-v3: Transcribe speech to text.
• Nemo_asr: Transcribe speech to text.

Speaker Diarization

• Pyannote 3.1: Identify and separate speakers in an audio file.

Voice Activity Detection (VAD)

• Silero VAD: Detect speech segments in audio files.

Text-to-Image Generation

• Stable Diffusion Medium-3: Generate images from text prompts.

Transformers-based Models

• llama2, llama3, llama3_1
• Mistralv2, Mistralv3
• Phi3.5 Mini
• AgentLM 7b

Quantized Models

• LLaMA 2, LLaMA 3, LLaMA 3.1
• Mistral v2, Mistral v3
• Phi3.5 Mini
• AgentLM 7b

Installation

To install the library and its dependencies, use the following command:

pip install -r requirements.txt

Usage Examples

1. Transformers

Transformers models are versatile and can be used for various NLP tasks. Here's an example using the LLaMA 3 model

from parent.factory import ModelFactory
model = ModelFactory.get_model("agentlm")  # No need to specify model_path
params = ModelFactory.load_params_from_json('parameters.json')
model.set_params(**params)
response = model.generate(
     prompt="What is Artificial Intelligence?",
     system_prompt="Answer in German."
 )
print(response)

Similarly use following string for other models of transformers:

• agentlm
• Phi3_5
• llama2
• llama3
• llama3_1
• Mistralv2
• Mistralv3

2. FastPitch (Text-to-Speech)

FastPitch is used for generating speech from text:

from parent.factory import ModelFactory
model = ModelFactory.get_model("fastpitch")

response = model.generate(text="Hello, this is Hasan Maqsood",output_path="Hasan.wav")

3. Voice Activity Detection (VAD)

Silero VAD is used for detecting speech timestamps in audio files:

from parent.factory import ModelFactory
model = ModelFactory.get_model("silero_vad")

response = model.generate("Youtube.wav")
print("Speech Timestamps:", response)

4. Speaker Diarization

Pyannote is used for speaker diarization:

from parent.factory import ModelFactory
model = ModelFactory.get_model("pyannote",use_auth_token="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")
response = model.generate("Hasan.wav", visualize =True)

5. Automatic Speech Recognition (Speech-To-Text)

Nemo ASR is used for transcribing audio to text:

from parent.factory import ModelFactory
model = ModelFactory.get_model("nemo_asr")
response = model.generate(audio_files=["Hasan.wav"])
print(response)

6. Distil/ Openai whsiper (Speech-To-Text)

Distil-whisper is used for transcribing audio to text:

from parent.factory import ModelFactory
model = ModelFactory.get_model("whisper")
response = model.generate("Youtube.wav")
print("Transcription:", response)

Openai-whisper is also used for transcribing audio to text:

 from parent.factory import ModelFactory
 model = ModelFactory.get_model("openai-whisper")
 response = model.generate("Youtube.wav")
 print("Transcription:", response)

7. Text-to-Image Generation

Stable Diffusion is used for generating images from text prompts:

from parent.factory import ModelFactory
model = ModelFactory.get_model("sd_medium3")
response = model.generate(prompt ="House")
image_path = "new_house.png"
response.save(image_path)