音频｜speech-2.8-hd 高清语音合成

MiniMax's speech-2.8-hd is their high-definition text-to-speech, and this node is a remarkably complete wrapper around it - 61 voices, per-language pronunciation boosting, speed/pitch/volume, emotion tags, a tone dictionary for name pronunciation, even a voice-blending JSON input. If you've ever fought a TTS node that gives you one model, one voice, and one rate, this is the opposite problem: the options are the feature. It's aimed at the pack's core audience of Chinese e-commerce content producers, but the voice list and language controls make it work for English narration too.

The mechanism under the hood is MiniMax's T2A service reached through Tikpan's relay (/minimax/v1/t2a_v2 synchronous, or the async variant for long text), and the node does a couple of smart things for you. The default return mode is hex - the audio comes back inline as hex and gets decoded locally into a proper audio file, which is more reliable than chasing a hosted URL. And it defaults 复用本地缓存 to on: identical text plus identical params hit a local cache and return the saved audio instead of re-billing you. That last one is genuinely considerate - failed or repeated runs are how people accidentally burn TTS credit.

The inputs that matter (and the syntax worth knowing)

• 合成文本 - the text, and it supports MiniMax's control syntax: <#0.5#> inserts a 0.5-second pause, and (laughs), (sighs), (coughs) add sound effects. This is the single best lever for making TTS sound like a person instead of a robot.
• 音色 - 61 voices, each labeled with gender/age/style in Chinese. The default is a Mandarin executive male voice.
• 语速 / 音量 / 音调 - rate and level multipliers, plus semitone pitch.
• 情绪 - happy/sad/angry etc.; leave at default to skip.
• 语言增强 - boost a specific language's pronunciation, auto otherwise.
• 采样率 / 比特率 / 音频格式 / 声道数 - output quality knobs; mp3/32000/128000 mono is a fine starting point.
• 开启字幕 - returns a subtitle timeline alongside the audio if you want to burn captions.
• 自定义voice_id, 音色混合_JSON, 发音字典, 声音效果 - the advanced lane: cloned/designed voices, voice blending with weights, per-word tone dict, and effects like spacious_echo.

Outputs are 🎧_音频流 (AUDIO), 📁_音频路径, 🔗_音频链接, 🆔_任务ID, 🆔_文件ID, 💰_计费字符数, and 📋_状态日志 - wire the AUDIO stream into a save/preview node and keep 计费字符数 handy for cost tracking.

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, type a sentence with a <#0.5#> pause and a (laughs), run. Nothing downloads; the model is server-side. Then start exploring voices - that's where most of the fun is.

Common issues

The pack-wide 401/402/429 family applies. TTS-specific: check 计费字符数 before long re-runs (repeated long text is how you discover the pricing), and keep 复用本地缓存 on so identical re-runs don't re-bill. If a sync call times out on long text, switch 调用方式 to async and collect via the task ID. If 🎧_音频流 won't preview but 📁_音频路径 has a file, the render succeeded - the preview gap is local. And the 校验HTTPS证书 toggle defaults to on here, so local-proxy setups that trip SSL should look there first.