关于 文生语音 模型
在 Modellix 探索 9 个生产可用的文生语音 AI 模型，对比模型能力、在线试用，并通过统一 API 快速完成集成。
全部 文生语音 模型
alibaba/qwen-audio-3.0-tts-flash
[Core Function] Qwen-Audio 3.0 TTS Flash is Alibaba's low-latency Qwen-Audio text-to-speech model on the same SpeechSynthesizer endpoint family as CosyVoice. [Strengths] Fast synthesis with voice, format, sample-rate, prosody, SSML, instruction, language_hint, and AIGC watermark controls; system voices include longanhuan_v3.6, longjielidou_v3.6, loongeva_v3.6, and loongjohn (see the Qwen-Audio-TTS voice list). [Best For] Voice assistants, interactive prompts, short announcements, multilingual product flows, and latency-sensitive batch TTS using Qwen-Audio voices. [Limitations] Does not expose hot_fix or Markdown filtering. Do NOT mix Plus-only voices (e.g. longanlingxin) with Flash. Use Plus when maximum narration quality matters more than turnaround time. [Routing] Choose Flash when speed matters most. Choose qwen-audio-3.0-tts-plus for premium narration quality.
alibaba/qwen-audio-3.0-tts-plus
[Core Function] Qwen-Audio 3.0 TTS Plus is Alibaba's high-quality Qwen-Audio text-to-speech model on the same SpeechSynthesizer endpoint family as CosyVoice. [Strengths] Natural speech synthesis with voice, format, sample-rate, prosody, SSML, instruction, language_hint, and AIGC watermark controls; system voices include longanlingxin and longanlufeng (see the Qwen-Audio-TTS voice list). [Best For] Premium narration, brand voiceovers, multilingual product audio, and quality-sensitive batch TTS when Qwen-Audio voices are preferred. [Limitations] Does not expose hot_fix or Markdown filtering. Do NOT use for real-time streaming or word-level timestamps. Do NOT mix Flash-only voices (e.g. longanhuan_v3.6) with Plus. [Routing] Choose Plus when speech quality is the priority. Choose qwen-audio-3.0-tts-flash when lower latency matters more.
xai/grok-voice-tts
[Core Function] Grok Voice TTS converts text into natural spoken audio with expressive voices and optional speech tags embedded in the text. [Strengths] Supports 20+ languages (plus auto-detect), 26 built-in voices, multiple codecs (mp3/wav/pcm/mulaw/alaw), and custom voice IDs. [Best For] Product voiceovers, IVR/telephony (mulaw/alaw), and multilingual narration. [Limitations] Do NOT exceed 15,000 characters per request. Audio is returned on the completed task after polling. [Routing] Use this model for xAI Grok Voice quality or custom cloned voices. [Built-in voices] Original: eve (default), ara, leo, rex, sal. Flagship: altair, atlas, carina, castor, celeste, cosmo, helios, helix, iris, kepler, lumen, luna, lux, naksh, orion, perseus, rigel, sirius, ursa, zagan, zenith — case-insensitive; custom voice IDs are also accepted as voice_id.
minimax/speech-2.8-turbo
[Core Function] MiniMax Speech 2.8 Turbo is a lower-latency text-to-speech model with the same control surface as Speech 2.8 HD, including paralinguistic tags such as (laughs). [Strengths] Faster and more cost-efficient synthesis while retaining prosody, timbre mix, pronunciation, subtitle, and audio-format controls. [Best For] Highly recommended for: interactive assistants, high-volume TTS batches, cost-sensitive voiceovers, quick narration drafts, and latency-sensitive product prompts. [Limitations] Do NOT use this for streaming or real-time partial audio. Do NOT request hex output or emotion whisper. Do NOT send both voice_id and timbre_weights. Prefer speech-2.8-hd if maximum audio quality is required. Audio URLs expire in about 24 hours. [Routing] Choose speech-2.8-turbo when the user emphasizes speed, cost, or throughput. Choose speech-2.8-hd for premium quality or highly expressive delivery.
minimax/speech-2.8-hd
[Core Function] MiniMax Speech 2.8 HD is a high-quality text-to-speech model that converts text into natural spoken audio, including expressive paralinguistic cues such as (laughs) and (sighs). [Strengths] Strong narration quality, stable prosody controls (speed, volume, pitch, emotion), optional timbre mixing, pronunciation overrides, subtitles, and flexible audio formats. [Best For] Highly recommended for: brand voiceovers, audiobook or long-form narration, marketing clips, multilingual delivery with language_boost, and expressive character speech. [Limitations] Do NOT use this for streaming or real-time partial audio. Do NOT request hex output or emotion whisper. Do NOT send both voice_id and timbre_weights. Audio URLs expire in about 24 hours. [Routing] Choose speech-2.8-hd when quality or expressive delivery matters most. Choose speech-2.8-turbo when the user prioritizes lower latency or cost.
google/gemini-3.1-flash-tts
[Core Function] Gemini 3.1 Flash TTS is Google's low-latency, controllable text-to-speech model. [Strengths] Single-speaker and two-speaker dialogue, 30 prebuilt voices, 70+ languages via language_code, and expressive delivery through style prompts plus inline audio tags such as [whispers], [slow], [fast], and [laughs]. Output is WAV (24 kHz mono PCM). [Best For] Voiceovers, virtual presenters, audiobook narration, multilingual speech, podcast-style scripts, and two-person dialogue. [Limitations] Do NOT use for lip-syncing existing video, music or sound-effect generation, image or audio inputs, MP3/OGG export, or API-level speed, volume, encoding, or sample-rate controls. Combined prompt and text must stay within 8,000 bytes. [Routing] Route here for controllable Gemini TTS from text only; for video with native audio use Veo; for talking-head lip sync from a portrait plus audio use Kling Avatar or SkyReels avatar models.
alibaba/cosyvoice-design
[Core Function] CosyVoice Design creates a temporary voice from a natural-language voice_prompt and synthesizes speech in one async request; only the final audio is returned. [Strengths] No voice enrollment management; language_hint (zh/en) applies to both design and synthesis; same prosody and format controls as CosyVoice TTS. [Best For] One-off designed speech without storing enrolled voices. [Limitations] Do NOT use this to obtain a reusable voice library entry; the designed voice is temporary and is not returned. model must be cosyvoice-v3.5-plus or cosyvoice-v3.5-flash; voice_prompt and text are required.
alibaba/cosyvoice-v3-flash
[Core Function] CosyVoice v3 Flash is Alibaba's low-latency text-to-speech model. [Strengths] Rich system voice catalog, fixed-format instruction on Instruct-capable system voices, SSML, hot_fix, AIGC watermark, Markdown filter (cloned voices only), and multiple audio formats with faster turnaround than Plus. [Best For] Voice assistants, interactive prompts, IVR, short announcements, dialect system voices, and latency-sensitive batch TTS. [Limitations] Do NOT use when maximum speech quality or long-form audiobook fidelity is the priority (use Plus). System-voice instruction must follow CosyVoice voice-list fixed Chinese formats. [Routing] Choose Flash when speed or a richer system-voice catalog matters most. Choose Plus for premium narration quality.
alibaba/cosyvoice-v3-plus
[Core Function] CosyVoice v3 Plus is Alibaba's high-quality text-to-speech model. [Strengths] System voices (e.g. longanyang, longanhuan), SSML and LaTeX input, hot_fix pronunciation correction, AIGC watermark, and output in mp3, pcm, wav, or opus. System-voice instruction must use the fixed Chinese formats in the CosyVoice voice list. [Best For] Brand voiceovers, audiobooks, high-quality narration, marketing clips, and scenarios where speech quality matters more than minimum latency. [Limitations] Do NOT use for real-time streaming or word-level timestamps. Fewer system voices than Flash. [Routing] Choose Plus when quality or narration fidelity matters most. Choose cosyvoice-v3-flash for lower latency or a richer system-voice catalog.
