repository·develop·Indexed 11 days ago
https://github.com/paddlepaddle/paddlespeechAn open-source toolkit on the PaddlePaddle platform for speech and audio tasks, providing state-of-the-art models for recognition, synthesis, and translation. It includes Audiotools for signal processing, support for various speech corpora like AISHELL-3 and GigaSpeech, and features such as Audio Content Search (ACS) for retrieving keyword timestamps from transcribed speech.
The Speech Server demo provides an implementation for starting a voice service and accessing it. Users can interact with the service either via command-line tools (paddlespeech_server and paddlespeech_client) or by writing a few lines of Python code.
For detailed service interface definitions, refer to the PaddleSpeech Server RESTful API documentation.
Paddle Speech Demo is a showcase project designed to help users get started with PaddleSpeech and build their own applications. It integrates speech interaction via PaddleSpeech, dialogue and information extraction via PaddleNLP, and a web frontend based on Vue3.
Core Functionalities:
Via main.py:
Via vc.py:
The TAL_CSASR dataset consists of English class audio featuring mixed Chinese and English speech. It is designed for tasks involving intra-sentence and inter-sentence code-switching. The dataset contains over 100 speakers (specifically 200+) and focuses on audio from English teachers' instruction.
Key Dataset Specifications:
.wav files..txt files (for test sets).The Speech Service demo provides an implementation for starting an offline speech service and accessing it. Users can interact with the service using single commands via paddlespeech_server and paddlespeech_client, or by writing a few lines of Python code.
For detailed interface definitions, refer to the PaddleSpeech Server RESTful API documentation.
This example provides a complete pipeline for training Transformer or Conformer-based Automatic Speech Recognition (ASR) models using the Librispeech dataset. It utilizes Kaldi functions and is managed via a central run.sh script.
Prerequisites:
--stage and --stop_stage flags.# Run stages 2 and 3 only
bash run.sh --stage 2 --stop_stage 3
# Run only stage 0
bash run.sh --stage 0 --stop_stage 0This example demonstrates how to train a VITS (Variational Inference with adversarial learning for end-to-end Text-to-Speech) model using the AISHELL-3 dataset. The workflow is designed to enable Voice Cloning by leveraging a speaker encoder trained via Speaker Verification (GE2E).
Key components:
The general pipeline involves: Speaker Encoder training $\rightarrow$ Data Preprocessing (generating embeddings) $\rightarrow$ Model Training $\rightarrow$ Waveform Synthesis $\rightarrow$ Voice Cloning Inference.
Parakeet is a text-to-speech (TTS) toolkit built on PaddlePaddle dynamic graph. It provides a standardized pipeline for TTS development, including data preprocessing, module sharing, model configuration, training, and synthesis.
The toolkit is organized into three main components:
Additionally, Parakeet supports Voice Cloning using transfer learning and GE2E (Generalized End-to-End Loss).
paddlespeech.s2t.frontend package provides the frontend components for Speech-to-Text (S2T) tasks. It is organized into subpackages for data augmentation and feature extraction, and submodules for handling audio processing, normalization, speech-specific logic, and general utilities.AISHELL-3 is a large-scale, high-fidelity multi-speaker Mandarin speech corpus designed for training multi-speaker Text-to-Speech (TTS) systems.
Key Characteristics:
Usage Note: This database is free for academic research only. Commercial use requires explicit permission.
The Audio Searching demo demonstrates how to build an audio similarity retrieval system capable of searching through large-scale audio datasets (e.g., hundreds of millions of tracks).
The demo is pre-configured to use the CN-Celeb dataset (650,000+ audio entries, 3,000 speakers), but it can be adjusted to use other datasets such as Librispeech, VoxCeleb, UrbanSound, GloVe, or MNIST.