arkennemasis fal · MiniMax Speech 2.6 HD · $0.1/1k chars

This is the generation where MiniMax TTS stopped being "good voice, no control." Three things arrived at once: real pause markers you type into the script, a built-in loudness normalization block, and a text budget that jumped from 5,000 to 10,000 characters. Still $0.10 per 1,000 characters.

What changed, and why you'd care

Pause markers. The field is called prompt on this generation (not text), and it accepts <#x#> where x is the pause in seconds, 0.01 to 99.99 to two decimals. You place them between speakable segments, and you can't use two consecutively. That's the difference between a narration that pauses where the writing does and one that pauses where the model felt like it.

Loudness normalization, on by default: target loudness −18 LUFS, target range 8 LU, target peak −0.5 dBTP - each of those a widget. If you've ever hand-compressed ten voice-over clips to get them to sit at the same level in an edit, you know what this is worth. It's also the reason a batch of clips from this node will cut together without a gain pass.

10,000 characters per call. Same rate, double the room - a full long-form script in one submit, no stitching.

How it works

Standard fal plumbing as the pack does it everywhere: FAL_KEY read from the .env beside the launcher, input validation, cost estimate from the character count against max_cost_usd, one billed submit, polling with a status label on the node, and a download into output/fal/fal-ai_minimax_speech-2.6-hd/. Nothing gets uploaded, so nothing before the submit can bill you. MiniMax's voice models are closed and hosted - there's no local equivalent, and that's the point: MiniMax positions this line against ElevenLabs, and the reason to pay is multi-language reliability that local TTS still doesn't quite match.

Inputs

• prompt - your script, 1 to 10,000 characters, with <#x#> pauses. This box sets the price.
• voice_setting_voice_id - free text, default Wise_Woman, MiniMax's predefined IDs. Empty leaves it to the model.
• voice_setting_emotion - (not set) or happy, sad, angry, fearful, disgusted, surprised, neutral.
• language_boost - the explicit language picker (with auto). Set it; this is what stops your German script being read with English instincts.
• voice_setting_speed (0.5–2.0), voice_setting_pitch (−12 to 12), voice_setting_vol (0.01–10), voice_setting_english_normalization (numbers and units read properly, slightly slower).
• Normalization: normalization_setting_enabled (default on), normalization_setting_target_loudness (−70 to −10 LUFS), normalization_setting_target_range (0–20 LU), normalization_setting_target_peak (−3 to 0 dBTP).
• 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), output_format (url or hex; leave it).

Optional: max_cost_usd ($20 default, 0 = off - the node refuses before sending when the estimate is over it), max_concurrent (default 1), reuse_identical_run (identical text returns the file you already paid for, restart included), extra_json (JSON merged into the request last).

Outputs: audio (waveform + sample rate, playable on the node), audio_path, duration_ms - the measured read, which you feed into Narration Length to size shots or Word Timings to time captions - 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 restart. And add the key:

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

in the .env next to run_nvidia_gpu.bat. It appears under arkennemasis/fal/Audio/MiniMax.

Using it well

Write pauses deliberately: <#0.6#> between clauses, <#1.2#> between paragraphs reads far more natural than a comma-and-hope. Leave english_normalization on for anything with numbers in it, and set language_boost even for English.

If you're targeting a platform rather than a broadcast mix, nudge target_loudness up toward the −16 to −14 LUFS range - most streaming platforms normalize around there, and delivering at −18 just means everything else in your edit is louder. And remember the reuse behaviour: re-running identical text returns the previous file for free, which is thrifty right up until it confuses you.