音频｜speech-2.8-turbo 极速语音合成

MiniMax's speech-2.8-turbo, which is the HD node's faster sibling: same 61 voices, same control syntax, same enormous parameter surface, tuned for latency instead of max audio quality. If your workflow is batch-generating hundreds of narration clips, or you're feeding TTS into a pipeline where the generation speed is the wall you keep hitting, this is the one to reach for. The tradeoff is audible but small - turbo voices sound a touch less produced than HD, and for short-form social audio most listeners won't tell the difference. The pack's own description is honest about it: same core parameters as HD, built for speed.

Mechanically it's the same MiniMax T2A relay - /minimax/v1/t2a_v2 (sync) or async for long text - with the same defaults that make the HD node nice to use: hex inline return, local caching to avoid double-billing on repeated identical calls, and the full voice/emotion/format surface preserved. The genuinely useful thing is that both nodes share a parameter vocabulary, so you can A/B the two models by swapping nodes and copying your inputs across - no re-learning.

The inputs that matter

• 合成文本 - the text, with <#0.5#> pause syntax and (laughs)-style sound-effect tags. Same as HD; this is the single highest-leverage field in the node.
• 音色 - the shared 61-voice list.
• 语速 / 音量 / 音调 - rate, level, semitone pitch.
• 情绪 - happy/sad/angry tags; default skips.
• 语言增强 - boost a specific language, auto to autodetect.
• 采样率 / 比特率 / 音频格式 / 声道数 - output encoding; mp3/mono is the sensible default for short-form.
• 调用方式 - sync for a single short clip, async for long text or batches (collect via the task ID afterward).

Advanced lane, shared with HD: 自定义voice_id, 音色混合_JSON (blend voices with weights), 发音字典 (per-word pronunciation), 声音效果 (e.g. spacious_echo), and 开启字幕 for a subtitle timeline.

Outputs mirror HD exactly: 🎧_音频流 (AUDIO), 📁_音频路径, 🔗_音频链接, 🆔_任务ID, 🆔_文件ID, 💰_计费字符数, 📋_状态日志.

Install and first run

cd ComfyUI/custom_nodes
git clone https://github.com/htrert/ComfyUI-Tikpan-Pro

Restart (or Manager → "Tikpan"), paste your sk- key, drop in a short sentence, run. Nothing to download - server-side model. Once it works, copy your voice and syntax settings into the HD node once to see if the quality difference is worth the latency for your use case.

Common issues

Same family as the HD node: 401/402/429, the 计费字符数 reality check before long re-runs, and keep 复用本地缓存 on so identical calls don't re-bill. Long text in sync mode can time out - switch to async. Preview gaps are local (file's on disk if 📁_音频路径 has it). And the 校验HTTPS证书 toggle defaults to on; local-proxy SSL pain points there.