arkennemasis fal · MiniMax Speech 02 HD · $0.1/1k chars

Speech 02 HD is MiniMax's flagship TTS - the model line the lab launched in January 2025 and benchmarked, openly, against ElevenLabs. This node is the fal endpoint wrapped for ComfyUI, priced at $0.10 per 1,000 characters. Paste a script, get an MP3 with a waveform you can play on the node.

Why you'd use it instead of a local TTS

Local TTS got genuinely good - Chatterbox at ElevenLabs-adjacent quality, Kokoro for speed, and both run on your own card. The KB's honest split is that the API still wins the two jobs where reliability matters most: best-in-class multi-language output, and production-grade delivery where a mispronounced name is expensive. MiniMax's own positioning is exactly that lane. If you're narrating in one language with a generic voice and it's not a paid deliverable, use Chatterbox and keep your money. If you need Arabic and German from the same script, or a specific predefined voice, this is the node.

How it works

The usual pack plumbing: FAL_KEY from the .env beside your launcher, input validation, a cost estimate sized from the character count of your text, one billed submit to fal, polling, then download into output/fal/fal-ai_minimax_speech-02-hd/. Nothing before the submit costs anything - and there are no uploads here at all, so a rejected input is entirely free.

Inputs

• text - note the field name: text, not prompt, and capped at 5000 characters. The price is straight arithmetic on this box, so a 500-word narration (~2,600 characters) is about 26 cents.
• voice_setting_voice_id - a free-text box defaulting to Wise_Woman, one of MiniMax's predefined voice IDs. Empty leaves the choice to the model. This is a predefined-voice picker, not a clone slot.
• voice_setting_emotion - (not set), happy, sad, angry, fearful, disgusted, surprised, neutral. (not set) lets the model decide from your text, which is usually the right default.
• voice_setting_speed (0.5–2.0), voice_setting_pitch (−12 to 12), voice_setting_vol (0.01–10) - the small dials. Speed is the one people actually use; 1.05 or 1.1 often cleans up a narration read that drags.
• language_boost - an explicit language picker (English, Chinese, Japanese, German, Arabic and a long tail, plus auto). Worth setting: it tells the model what language it's reading, which is how you stop numbers and names from being read with the wrong accent.
• voice_setting_english_normalization - off by default; turn it on if your script is number-heavy and reads badly.
• audio_setting_format (mp3, pcm, flac), audio_setting_bitrate (32000–256000), audio_setting_sample_rate (8000–44100), audio_setting_channel (1 = mono, 2 = stereo).
• pronunciation_dict - a JSON box for text replacements, which is the escape hatch for a name the model keeps mangling. Empty leaves it out.
• output_format - url or hex, default hex, non-streaming only. Leave it as shipped.

Optional: max_cost_usd (default $20; the run halts before sending if the estimate is over it, 0 = off), max_concurrent (default 1), reuse_identical_run (identical text and settings return the file you already paid for, restart included), extra_json.

Outputs: audio (waveform + sample rate, playable on the node), audio_path, duration_ms - an integer you can wire into the pack's Narration Length or Word Timings to time shots and captions to the real read - and info.

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, and it's under arkennemasis/fal/Audio/MiniMax.

The tier question

HD at $0.10/1,000 characters versus Turbo at $0.06. Same model family, same inputs, same voice list. Turbo is the one for "does this script read well?" - you're paying 60% for a check. HD earns its price on the final render, where the difference in warmth and breathing is the thing people notice. Iterate cheap, deliver HD.