Two questions behind a waveform

A player handles a changing audio waveform; a script consists of characters and words. A speech model must connect them while deciding how long sounds last, how pitch moves and which speaker characteristics to retain.

Ask two separate questions: what represents the sound, and how does the model generate that representation? Discrete versus continuous mostly concerns the first. Autoregression and diffusion describe generation. They do not form a single ladder from old to new.

Qwen3-TTS organizes speech with discrete codes

The Qwen3-TTS 12Hz route uses a multi-codebook speech tokenizer to represent compressed audio as discrete codes, with autoregressive modeling generating the sequence for decoding. Think of them as machine-readable sound symbols, not letters or one symbol per word. See the technical report for the architecture.

The family separates reference-based cloning, preset speakers and described voices into Base, CustomVoice and VoiceDesign. For a creator, that changes which material to prepare. A reference recording belongs in a route that accepts audio; a written character description belongs in voice design first.

VoxCPM2 models continuous audio latents

VoxCPM2 uses continuous audio representations with hierarchical diffusion-autoregressive modeling. Tokenizer-free does not mean uncompressed: AudioVAE still represents and reconstructs audio. The distinction is avoiding an external discrete speech tokenizer as the modeling step. The official report explains this approach.

Different arrangements of inputs support design, cloning and style guidance within one model. This offers several ways to work with a description or recording. Continuous representation alone does not prove better narration: training data, language, reference material and runtime still affect what you hear.

OmniVoice combines discrete codes with diffusion

OmniVoice shows why discrete representation and diffusion are compatible. Its paper describes a diffusion-language-model-style discrete non-autoregressive architecture mapping text directly to multi-codebook acoustic tokens. It iteratively refines codes rather than relying on a conventional left-to-right sequence alone. This describes architecture, not a speed prediction for your Mac.

The project emphasizes broad language coverage and offers cloning and design. Upstream notes that design training uses Chinese and English data. Coverage and reliable description-based voice creation are separate questions, especially for less represented languages.

What architecture can explain

Architecture helps explain differences in reference inputs and generation controls, such as sampling settings or iterative steps. It cannot choose the best narrator for a novel by itself. Model size, runtime, quantization and the script also influence waiting time and delivery.

Audio encoding and weight quantization are different forms of compression. The first represents sound; the second changes how model numbers are stored and computed. A 4bit or 8bit model does not imply four-bit or eight-bit output audio. A higher output sample rate also cannot repair missing sentences or misplaced emphasis.

Compare on the same piece of work

Use the same script with long sentences, numbers, dialogue and a change of context. Check the words, then continuity and expression, and finally waiting and correction time. Where supported, keep the reference recording fixed across engines.

Sonano lets compatible reference-based models share saved designed and cloned voices within language constraints. Native Kokoro and CustomVoice presets use a separate route. Preview again after changing engines. Read how design, cloning and generation fit together, or check the model catalog. This article explains approaches; it is not a measured ranking.

Sources & further reading

Sources checked on November 3, 2026. Creative scenarios and listening examples are illustrative, not executed model tests. App capabilities depend on the current version.