Search Results for "real-time vocoder"

WhisperSpeech is an open-source text-to-speech system created by “inverting” OpenAI’s Whisper, reusing its strengths as a semantic audio model to generate speech instead of only transcribing it. The project aims to be for speech what Stable Diffusion is for images: powerful, hackable, and safe for commercial use, with code under Apache-2.0/MIT and models trained only on properly licensed data. Its architecture follows a token-based, multi-stage pipeline inspired by AudioLM and SPEAR-TTS:...

Parallel WaveGAN is an unofficial PyTorch implementation of several state-of-the-art non-autoregressive neural vocoders, centered on Parallel WaveGAN but also including MelGAN, Multiband-MelGAN, HiFi-GAN, and StyleMelGAN. Its main goal is to provide a real-time neural vocoder that can turn mel spectrograms into high-quality speech audio efficiently. The repository is designed to work hand-in-hand with ESPnet-TTS and NVIDIA Tacotron2-style front ends, so you can build complete TTS or singing voice synthesis pipelines. It includes a large collection of “Kaldi-style” recipes for many datasets such as LJSpeech, LibriTTS, VCTK, JSUT, CMU Arctic, and multiple singing voice corpora in Japanese, Mandarin, Korean, and more. ...

MockingBird is an open-source voice cloning and real-time speech generation toolkit that lets you clone a speaker’s voice from a short audio sample (reportedly as little as 5 seconds) and then synthesize arbitrary speech in that voice. It builds on deep-learning based TTS / voice-cloning technology (in the lineage of projects such as Real-Time-Voice-Cloning), but extends it with support for Mandarin Chinese and multiple Chinese speech datasets — broadening its applicability beyond English. ...

VoiceFixer is a machine-learning framework for “speech restoration”: given a degraded or distorted audio recording — with noise, clipping, low sampling rate, reverberation, or other artifacts — it attempts to recover high-fidelity, clean speech. The architecture works in two stages: first an analysis stage that tries to extract “clean” intermediate features from the noisy audio (e.g. removing noise, denoising, dereverberation, upsampling), and then a neural vocoder-based synthesis stage that...

...It offers a variety of architectures for text-to-speech, including classic and modern models such as Tacotron‑2, FastSpeech / FastSpeech2, and neural vocoders like MelGAN and Multiband‑MelGAN. Because it’s based on TensorFlow 2, it can leverage optimizations such as fake-quantization aware training and pruning — which allow models to run faster than real time and to be deployable on mobile or embedded platforms. The library supports multiple languages (English, French, Korean, Chinese, German, etc.) and is relatively easy to adapt to new languages. With integrated vocoder + mel-spectrogram generation pipelines, pre-trained models, and fairly flexible architecture, TensorFlowTTS is a great off-the-shelf and extensible TTS engine for applications ranging from voice assistants to content generation or accessibility tools.

...In experiments on LJSpeech, HiFi-GAN was shown to achieve mean opinion scores close to human recordings while synthesizing 22.05 kHz audio up to ~168× faster than real time on an NVIDIA V100 GPU. A smaller configuration trades a bit of quality for even higher speed and can run more than 13× faster than real time on CPU, making it suitable for deployment scenarios without powerful GPUs.

...It includes directory layouts and logging directories for multiple datasets such as LJSpeech and M-AILABS en_US/en_UK, making it easier to adapt to new English corpora. Separate log trees track mel-spectrograms, attention plots, evaluation audio, and vocoder outputs, so you can inspect how alignment and audio quality evolve over time.