MiniMax Text to Speech

The MiniMax Text to Speech node is the reason most people install this pack. It turns up to 5,000 characters of text into natural-sounding, MP3 speech using MiniMax's commercial TTS models - the same voices that show up as the benchmark next to ElevenLabs and GPT-4o-audio in every voice-model comparison thread. The quality is genuinely good, the emotion and speed controls work, and you get it as a node in your graph instead of a web app you have to tab away to.

How it works

The node POSTs to https://api.minimaxi.chat/v1/t2a_v2?GroupId=... with your text, a voice_setting block (voice, speed, volume, pitch, optional emotion), and an audio_setting (32kHz, 128kbps MP3, mono - fixed). The response comes back as hex-encoded audio (default output_format: hex) or a URL, and the node writes a timestamped .mp3 to your ComfyUI output folder. If subtitle_enable is on, it also fetches a JSON subtitle file with sentence-level timestamps.

One honest quirk: the seed input exists and does... not what a diffusion user expects. The node explicitly does not send it to the API - it's only used for ComfyUI's execution/caching control, and it generates a random value internally when you leave it at 0. Don't expect a fixed seed to give you bit-identical audio. This is a cloud API, not a local sampler.

The inputs that matter

The two that get people stuck:

• api_key and group_id - both required. TTS (and Voice Cloning) need the Group ID from your MiniMax console, not just the key. Get both from MiniMax's open platform, and expect a "Group ID must be provided" error if you skip the second one.

The ones you'll actually fiddle with:

• text - what to say, up to 5,000 characters.
• model - speech-2.5-hd-preview (default), speech-02-hd/speech-02-turbo, or the older speech-01-hd/speech-01-turbo. The .hd models sound better; the -turbo ones are faster and cheaper. Start with the default.
• voice_id - a predefined voice from MiniMax's library. If you leave it empty, the node uses whatever falls out of the custom_voice_id path below - so don't.
• custom_voice_id - this is the one you wire from the Voice Cloning or Voice Design node. If set, it overrides voice_id.
• speed (0.5–2.0), volume (0.1–10.0), pitch (−12 to +12) - the standard mix desk. Pitch in semitones, roughly.
• emotion - happy, sad, angry, fearful, disgusted, surprised, or neutral. Only applied if you pick one; empty means the model's default delivery.
• language_boost - auto by default, or pick the language if the model is mangling accents.

The nice-to-haves: subtitle_enable for a timestamped JSON, filename_prefix for tidy outputs, and output_format (hex default, or url for a 24-hour hosted link).

Outputs

Three strings: audio_path (the saved MP3 - wire it into Preview Audio), subtitle_path (the JSON, if enabled), and audio_url (only populated in url mode).

Common issues

Missing group_id is the top failure. Past that, the API errors are raised as readable messages (auth, balance, rate limit). And remember this is a paid cloud call - every generation burns a little API credit, so preview with short text before you render 5,000 characters.

Install

Part of the ComfyUI-JM-MiniMax-API pack:

cd ComfyUI/custom_nodes
git clone https://github.com/synthetai/ComfyUI-JM-MiniMax-API
pip install -r requirements.txt

Or search "ComfyUI-JM-MiniMax-API" in ComfyUI Manager and restart. Just requests and Pillow - no GPU, no local models. The whole pack is a thin wrapper over MiniMax's cloud, which is exactly why it installs in ten seconds.