arkennemasis fal · MiniMax Speech 2.8 Turbo · $0.06/1k chars

If you're building anything with a voice-over in it - a narrated short, an explainer, a talking head - the speech is the part you'll re-run fifty times as the script changes. That's exactly why the price matters. MiniMax's Speech 2.8 Turbo bills at $0.06 per 1,000 characters on fal, and 1,000 characters is roughly 150 words, roughly a minute of talking. A three-minute narration costs about eighteen cents. You can re-roll the script all afternoon and never watch a meter.

This is a hosted node. Nothing here runs on your GPU, and MiniMax's whole voice line is closed and API-first - Hailuo's TTS is the product they benchmark against ElevenLabs. That's the trade: zero VRAM, zero model download, but your text goes to fal and comes back as audio.

Why you'd route out for this one

The knowledge base's audio chapter lands on internal TTS is genuinely good now - Chatterbox, Kokoro, F5, and the pack's own local Qwen3-TTS node. Still, the API path keeps one job: production narration where a mispronunciation is expensive and you don't want to babysit a model. It also costs no VRAM, which matters when a 14B video model is sitting on your card.

How the node actually works

One submit to fal, one billed call. Before that: the node reads FAL_KEY from the .env in your ComfyUI install, estimates the cost from your prompt's character count, and stops cold if the estimate is over max_cost_usd (default $20 - yes, that's a lot of speech). Nothing is uploaded before the estimate passes, so a failure up to that point never bills. The finished audio lands in output/fal/fal-ai_minimax_speech-2.8-turbo/ and plays on the node.

There's no key box anywhere in this pack, on purpose - a key typed into a node gets saved into every workflow JSON you share.

The settings worth touching

The text box is the whole job, but it has two bits of syntax that surprise people. <#1.5#> inserts a pause of that many seconds (0.01–99.99), and interjection tags are performed rather than read: (laughs), (sighs), (coughs), (clears throat), (gasps), (sniffs), (groans), (yawns). On a short line, that's the difference between a robot and a person.

Then the handful that actually matter:

• voice_setting_voice_id - a plain text string, default Wise_Woman. Note string, not a combo: the pack's Voice Clone and Voice Design nodes return custom_voice_id as a STRING, so you can wire one straight into the other.
• voice_setting_english_normalization - turn it on the moment your script contains $1,299, a date, or an acronym. Slight latency cost, far fewer "one thousand two hundred ninety nine dollars" style misses.
• language_boost - for anything non-English, say so. Left at (not set), the model guesses.
• voice_setting_emotion - happy / sad / angry / fearful / disgusted / surprised / neutral, or leave it to the model.
• audio_setting_* - mp3, 32 kHz, mono by default. Switch to flac and 44100 if this is going under music.
• normalization_setting_enabled - on by default at −18 LUFS with a −0.5 dBTP peak, which is exactly what you want when you're stacking narration against a music bed.

voice_modify_pitch / intensity / timbre (−100 to 100) are the "same person, angrier" knobs - shaping, not casting. And pronunciation_dict takes JSON text replacement for names the model keeps mangling.

Outputs

audio is a real ComfyUI AUDIO value with a player on the node - wire it into a sync node or a save node. audio_path is the same file as a string. duration_ms is the one people forget: it's how you size a shot before you spend on video. Wire it into the pack's Narration Length node and every clip outlasts its own voice-over.

Install

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

Restart ComfyUI. Or in ComfyUI-Manager, search arkennemasis (pack title comfyui-arkennemasis) and install from there. The pack's requirements.txt is just httpx, and that's for the Codex nodes - the fal provider talks to fal.ai with the standard library, so this path adds nothing to install.

Then put your key where the nodes look for it - a .env file next to run_nvidia_gpu.bat (or the ComfyUI folder):

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

Where people get burned

• The node does nothing, or 401s. The key isn't in the .env the pack reads. There's no key field on the node - that's deliberate, not an oversight.
• Twenty fal nodes, one at a time. max_concurrent defaults to 1 across the whole run, so nodes sit at "waiting for its turn". Raise it on the nodes you want in parallel, up to 32.
• It re-read your numbers as digits, or drifted into the wrong accent. english_normalization, language_boost.
• You paid twice for the same line. Don't: reuse_identical_run is on by default and returns the same files for identical inputs without billing, even after a restart. Turn it off when you want a fresh take.
• You want a different read, not a different voice. Change voice_setting_emotion and voice_modify_* before you go hunting for another voice ID.