Search Results for "semantic"

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: Whisper is used to produce semantic tokens, EnCodec compresses the waveform into acoustic tokens, and Vocos reconstructs high-fidelity audio from those tokens. ...

...Unlike traditional TTS systems, it excels in scalability, speaker consistency, and natural turn-taking for up to 90 minutes of continuous speech with as many as four distinct speakers. A key innovation is its use of continuous acoustic and semantic speech tokenizers operating at an ultra-low frame rate of 7.5 Hz, enabling high audio fidelity with efficient processing of long sequences. The model integrates a Qwen2.5-based large language model with a diffusion head to produce realistic acoustic details and capture conversational context. Training involved curriculum learning with increasing sequence lengths up to 65K tokens, allowing VibeVoice to handle very long dialogues effectively. ...

Step-Audio is a unified, open-source framework aimed at building intelligent speech systems that combine both comprehension and generation: it integrates large language models (LLMs) with speech input/output to handle not only semantic understanding but also rich vocal characteristics like tone, style, dialect, emotion, and prosody. The design moves beyond traditional separate-component pipelines (ASR → text model → TTS), instead offering a multimodal model that ingests speech or audio and produces speech accordingly, enabling natural dialogue, voice cloning, and expressive speech synthesis. ...

...Rather than treating audio editing as low-level waveform manipulation, this model converts speech into a sequence of discrete “audio tokens” (via a dual-codebook tokenizer) — combining a linguistic token stream and a semantic (prosody/emotion/style) token stream — thereby abstracting audio editing into high-level token operations. This allows users to modify not only what is said (the text) but also how it's said: emotion, tone, speaking style, prosody, accent, even paralinguistic cues. Because the model is trained with a “large-margin learning” objective over many synthesized and natural speech samples, it gains robust control over expressive attributes, and can perform iterative editing: e.g. you could record a line, then ask the model to “make it sadder,” “speak slower,” or “change accent to X.”