AI 文本至发言稿

Bark by Suno is a transformer-based text-to-audio model that can generate highly realistic, multilingual speech as well as other audio like music, background noise, and sound effects. It can produce nonverbal communications like laughing, sighing, and crying. Bark supports over 100 speaker presets and 13+ languages.

Bark Small is a distilled version of the Bark model that trades some audio quality for significantly faster inference speeds and lower memory requirements. It retains Bark's ability to generate speech with emotions, laughter, and multiple languages.

CosyVoice 2 by Alibaba's Tongyi Lab achieves human-comparable speech quality with extremely low latency, making it ideal for real-time applications. It uses a finite scalar quantization approach for streaming synthesis and supports zero-shot voice cloning, cross-lingual synthesis, and fine-grained emotion control. It outperforms many commercial TTS systems in subjective evaluations.

Dia by Nari Labs is a 1.6B parameter text-to-speech model designed specifically for generating multi-speaker dialogue. It can produce natural-sounding conversations between two speakers with appropriate turn-taking, prosody, and emotional expression. Dia is perfect for creating podcast-style content, audiobook dialogues, and interactive conversational AI.

Parler TTS is a text-to-speech model that uses natural language voice descriptions to control the generated speech. Instead of selecting from preset voices, you describe the voice you want (e.g., "a warm female voice with a slight British accent, speaking slowly and clearly") and Parler generates speech matching that description. This makes it uniquely flexible for creative applications.

Indic Parler TTS by AI4Bharat extends the Parler architecture to Indian languages, generating natural speech in Tamil, Bengali, Marathi, Gujarati, Kannada, Punjabi, Odia, Assamese, Hindi, Telugu, Malayalam and English. Like Parler, you describe the voice you want in plain language and the model matches it — no preset voices required. Trained on AI4Bharat speech corpora for authentic pronunciation and prosody across the Indian subcontinent.

KhanomTan TTS is an open Thai text-to-speech model built on the YourTTS multilingual architecture. Trained on CC0 and permissively-licensed Thai corpora (TSync) alongside several other languages, it delivers natural Thai speech with multiple speaker voices. A clean, commercially-usable option for Thai — the language most open TTS models only cover under non-commercial licenses.

IndexTTS-2 is an advanced text-to-speech system that excels at zero-shot voice synthesis with fine-grained emotion control. It can generate speech with specific emotional tones like happy, sad, angry, or fearful without requiring emotion-specific training data. The model uses emotion vectors to precisely control the emotional expression of generated speech.

Spark TTS by SparkAudio is a text-to-speech model that combines voice cloning with controllable emotion and speaking style. Using just 5 seconds of reference audio, it can clone a voice and then generate speech with different emotions, speeds, and styles while maintaining the cloned voice identity. Spark TTS uses a prompt-based control system.

GPT-SoVITS combines GPT-style language modeling with SoVITS (Singing Voice Inference via Translation and Synthesis) for powerful few-shot voice cloning. With as little as 5 seconds of reference audio, it can accurately clone a voice and generate new speech while preserving the speaker's unique characteristics. It excels at both speaking and singing voice synthesis.

Orpheus is a large-scale text-to-speech model that achieves human-level emotional expression. Trained on over 100,000 hours of diverse speech data, it excels at generating speech with natural emotions, emphasis, and speaking styles. Orpheus can produce speech that is virtually indistinguishable from human recordings.

Qwen3-TTS is a 1.7 billion parameter text-to-speech model from Alibaba's Qwen team. It supports two modes: preset voices with emotion control (9 speakers), and a unique voice design mode where you describe the voice you want in natural language. It covers 10 languages with high expressiveness and natural prosody.

VieNeu-TTS-v2 is a 300M parameter Vietnamese-first TTS model trained on 10,000+ hours of bilingual data. It supports seamless en-vi code-switching, 7 preset voices spanning Northern and Southern accents, and instant voice cloning from 3-5 seconds of reference audio. Runs entirely on CPU via GGUF Q4 inference + ONNX audio decoder — no GPU needed, generations finish in ~7 seconds. Built on a Qwen3 backbone.

Chatterbox Turbo by Resemble AI is a 350M parameter upgrade to Chatterbox, delivering up to 6x real-time speed with sub-200ms latency. It supports paralinguistic tags like [laugh], [cough], and [chuckle] directly in text. Includes Perth watermarking on all generated audio for provenance tracking.

VoxCPM 1.5 by OpenBMB is a novel tokenizer-free TTS model that operates in continuous space rather than discrete tokens. It produces high-fidelity 44.1kHz audio, supports zero-shot voice cloning from 3-10 seconds, and maintains consistency across paragraphs. Cross-language cloning lets you apply an English voice to Chinese speech and vice versa.

VibeVoice from Microsoft generates long-form speech up to 90 minutes with support for 4 simultaneous speakers, making it ideal for podcasts and dialogues. The Realtime 0.5B variant achieves ~300ms latency for interactive use. Supports speaker tags for multi-turn dialogue generation.

CosyVoice3 is the latest evolution from Alibaba's FunAudioLLM team. It features bi-streaming inference with ~150ms latency, instruction-based control for emotion/speed/volume, and improved speaker similarity for zero-shot cloning. Supports 9 languages plus 18 Chinese dialects. RL-tuned variant delivers state-of-the-art prosody.

NAMAA Saudi TTS is a Saudi Arabic fine-tune of Resemble AI's ChatterboxMultilingual. Trained by NAMAA Space on authentic Saudi-dialect speech, it produces natural Modern Standard Arabic and Saudi colloquial pronunciation that generic multilingual models cannot match. Inherits Chatterbox's zero-shot voice cloning and emotion control via reference audio prompts. The first open-weights Arabic TTS deployed on TTS.ai.

Darwin-TTS-1.7B-Cross by FINAL-Bench is a research variant of Qwen3-TTS-1.7B where 84 talker-FFN tensors (8.6%) are blended at α=3% with the matching tensors from Qwen3-1.7B-Base. The blend is built without retraining and produces noticeably crisper cross-lingual voice cloning across Korean, English, Japanese, and Chinese. Operates in zero-shot voice-clone mode (3 seconds reference audio).

MOSS-TTSD v1.0 from OpenMOSS is a 7B dialogue text-to-speech model that continues conversations from a short audio prompt. Supports up to 5 simultaneous speakers via [S1]/[S2] tags, zero-shot voice cloning from 3-10s reference audio, and up to 60 minutes of coherent multi-turn dialogue across 20 languages. Distinct from MOSS-TTS — TTSD is specialized for podcast/audiobook/dubbing workflows.