Gemini TTS Multi API

This is the party trick. The plain Gemini TTS node gives you one voice per call; this one renders a whole conversation in a single generation, with a different voice for every speaker. Podcast banter, audiobook dialogue, two characters bickering - you write the script with speaker labels, assign each speaker a voice, and the model produces one continuous audio file with the cast already separated. No splicing clips together, no crossfade hacks.

How it works

Same REST engine as its single-voice sibling, one crucial difference in the payload. Instead of a single prebuiltVoiceConfig, the node builds a multiSpeakerVoiceConfig with a speakerVoiceConfigs array - one entry per speaker, each mapping a speaker name to a voice. Your script uses those names as labels ("Speaker 1: …", "Speaker 2: …"), and the model renders each labelled line in the assigned voice. The response is raw PCM audio, decoded back into a standard ComfyUI AUDIO dict.

What you feed it

The inputs that do the work:

• text - the script. The default is a Chinese-language template with three sections: scene background, director's notes (style, pace, accent guidance), and the transcript - the exact lines with speaker labels. You can keep that structure and swap the transcript for your own dialogue. The names in the transcript must match your configured speaker names exactly.
• speaker_options - a list of {speaker, voiceName} entries. This comes from the pack's own helper nodes, not typed by hand: Gemini Speaker Options builds one speaker, Gemini Batch Speakers Options merges two arrays. So a two-speaker setup is Speaker Options → Batch Speakers → this node.
• model - the same combo as the single TTS node, default gemini-2.5-flash-preview-tts.

Outputs are AUDIO (the full dialogue) and STRING (token usage or the error JSON). There's a seed input that, like elsewhere in this pack, is accepted but not sent to the API - don't lean on it for reproducibility.

Setting up two voices

• One Gemini Speaker Options node per speaker: set speaker to the exact label used in the transcript (e.g. Speaker 1) and pick a voiceName (Zephyr, Puck, Charon…).
• Run both into Gemini Batch Speakers Options, then into this node's speaker_options.
• Keep transcript labels and speaker names in lockstep - a mismatch is the most common way this 400s.

Installing and the usual caveats

Same pack as everything above:

cd ComfyUI/custom_nodes
git clone https://github.com/ycyy/ComfyUI-YCYY-API

Copy config.json.example to config.json and add your key to the gemini-tts section. No models, no GPU, no extra deps - restart and look under YCYY/API.

The gotchas are inherited from single-voice TTS, plus one new one: duplicate speaker names are rejected, so every label must be unique, and every transcript label must have a matching speaker entry. The node also validates that speaker_options is a non-empty list - forget to wire the options in and you'll get an error before a single byte of audio comes back.

As always with this pack, failures don't throw - they return (None, error_json) out of the two outputs. Check the STRING, not the AUDIO. Multi-speaker mode is also a newer API surface than plain TTS, so if you start seeing 400s after a Google update, that's usually the model name churning in config rather than a problem with your script.