arkennemasis fal · MiniMax Preview Speech 2.5 HD · $0.1/1k chars

"Preview" in the endpoint name makes it sound like a work-in-progress. It isn't a downgrade in the way you'd assume - this is the Speech 2.5 generation on fal's preview route, priced identically to the Speech 02 HD node at $0.10 per 1,000 characters. The thing that actually separates it from its siblings is the language list.

What it is, and why the language list matters

MiniMax's voice line is the part of their catalogue they benchmark openly against ElevenLabs, and the practical reason to pay for TTS at all is the same everywhere: local models like Chatterbox and Kokoro are excellent in a single language and less reliable across many. This node exposes a language_boost picker that includes Persian, Filipino and Tamil on top of the usual Chinese/English/Japanese/Korean/European set - and those three are exactly the ones missing from the 2.6 and 2.8 node lists in this pack. If your narration is in Farsi or Tagalog, that's not a nice-to-have, that's the reason you're on this node.

How it works

Standard pack behaviour over fal's minimax/preview/speech-2.5-hd endpoint: read FAL_KEY from the .env beside your launcher, validate the text, estimate cost from the character count, submit one billed request, poll while the node shows status, then download into output/fal/fal-ai_minimax_preview_speech-2.5-hd/. No uploads, so nothing before the submit can cost you anything.

One small confusion to head off: fal's own description text for these preview endpoints still says "using the MiniMax Speech-02 HD model" - the marketing copy wasn't updated when the endpoint moved on. Ignore the blurb; the box list is what the model takes.

Inputs

• text - the field is text on this generation (the 2.6 and 2.8 nodes call it prompt), capped at 5000 characters. Your bill tracks this box: ~500 words is about 26 cents.
• voice_setting_voice_id - free text, default Wise_Woman, from MiniMax's predefined voice IDs. Empty leaves it to the model. This is not a cloning slot.
• voice_setting_emotion - (not set) plus happy, sad, angry, fearful, disgusted, surprised, neutral.
• language_boost - set it. An English script read with the wrong language hint is where "the model can't pronounce anything" complaints come from.
• voice_setting_speed (0.5–2.0), voice_setting_pitch (−12 to 12), voice_setting_vol (0.01–10), voice_setting_english_normalization (off by default).
• Output: audio_setting_format (mp3, pcm, flac), audio_setting_bitrate (32k–256k), audio_setting_sample_rate (8k–44.1k), audio_setting_channel (1 mono, 2 stereo), pronunciation_dict (JSON replacements for names the model butchers), output_format (url or hex, non-streaming only - keep the default).

Optional, same four as every node in the pack: max_cost_usd ($20 default; the run stops before sending when the estimate exceeds it, 0 = uncapped), max_concurrent (default 1 - one paid fal call at a time in a run), reuse_identical_run (default on: identical text and settings return your already-paid file, even after a restart), extra_json.

Outputs: audio (waveform and sample rate, with a player on the node), audio_path, duration_ms - wire that integer into the pack's Narration Length so each shot covers its own voice-over - and info (request id, estimate, saved files, fal's answer).

Install

cd ComfyUI/custom_nodes
git clone https://github.com/Hishamahmer/comfyui-arkennemasis
pip install httpx

Or ComfyUI Manager → Install via Git URL. Then:

FAL_KEY=your-key-from-fal.ai/dashboard/keys

in the .env next to run_nvidia_gpu.bat. Restart; the node is under arkennemasis/fal/Audio/MiniMax.

Notes

The 5000-character cap is a hard stop, and no pause markers exist on this generation - the <#x#> syntax showed up later. If you need a deliberate beat in the read, use a full stop and a new sentence, or splice it in your editor.

For a script you're still editing, run the Turbo sibling at $0.06 instead and keep this one for the take you'll ship. The inputs are identical, so switching is a copy-paste, not a rebuild.