音频｜Gemini 3.1 Flash TTS

Google's Gemini TTS models are closed and API-only, and this node is the relay path into them from a ComfyUI graph. TikpanGemini31FlashTTSNode wraps gemini-3.1-flash-tts-preview behind a Tikpan key, and the headline feature is the 31-voice preset picker - named voices like Kore (Firm, for business narration), Puck (Upbeat, short-video friendly), Charon (Informative), Fenrir (Excitable) - each annotated with style and use case, so you pick a vibe instead of guessing a voice ID. It's positioned as low-latency TTS for narration and 口播 (spoken ad copy) work, billed per input/output token rather than per character.

How it works

Text goes to the relay's Gemini TTS endpoint; the node handles the streaming response, writes the audio, and hands back the usual six outputs: 📁_音频路径, 🔗_音频链接, 💰_输入输出用量 (token usage - this is your billing receipt), 🧾_接口路径 (which endpoint protocol ran), 📋_状态日志, and 🎧_音频流 (AUDIO for downstream nodes). The 调用方式 field is the interesting lever: geminitts 原生 (native, most stable - default), gemini 原生, or openai 兼容 (most interoperable if you're swapping this node for an OpenAI-style consumer). Pick native and leave it.

The inputs that actually matter

• API_密钥 - sk- from tikpan.com.
• 合成文本 - text to speak; the tooltip notes it supports mixed Chinese/English and multiple languages.
• 音色 - 31 presets. Kore (default) is the business-narration workhorse; Puck is the short-video energy voice.
• 语气指令 - a natural-language delivery instruction ("像纪录片旁白一样说", "低声说") merged into the text. This is the expressive knob the Doubao node fakes with emotion tags - Gemini takes actual prose direction.
• 语言代码 - auto, or force zh-CN / en-US / ja-JP / ko-KR / yue-HK / fr-FR / de-DE / es-ES.
• 采样率 - 24000 Hz only, per the schema. Fine for narration.

Optional: 自定义voice_name (new official voices or relay aliases), 复用本地缓存 (default on - same text+params hits local cache and avoids a double token charge), 跳过错误 (silent clip on failure to keep batch flows alive).

Installing it

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

restart (or Manager → "Tikpan"). No models, no heavy deps.

Where people get burned

• Token billing is sneaky. It's billed on input and output tokens, and the 💰_输入输出用量 output exists so you can actually watch it. Long instructions + long text = real spend; the local cache helps on repeats.
• Cert verification defaults ON. Correct behavior, but a corporate proxy will throw cert errors until you toggle 校验HTTPS证书.
• Don't overuse the instruction field. 语气指令 is powerful but every token you add is billed - a short, specific delivery note beats a paragraph.
• Three protocols, one node. If a workflow that previously used an OpenAI-compatible consumer stops working after a swap, the 调用方式 setting is the first thing to check.

Verdict: the strongest pure-voice picker in the pack for expressive multilingual narration, with the caveat that token metering makes the per-character Doubao node cheaper for long Chinese scripts. Choose by what you're voicing, not by hype.