arkennemasis fal · MiniMax Speech 2.6 Turbo · $0.06/1k chars

The 2.6 feature set at the Turbo price. Six cents per 1,000 characters, prompt instead of text, <#x#> pause markers, the loudness normalization block, and a 10,000-character budget. If you produce a lot of narration and you're not chasing the last 4 cents of fidelity, this is the node that should be doing most of your work.

What it is

The pack's node over fal's minimax/speech-2.6-turbo. MiniMax's voice line is the piece of their catalogue they benchmark head-on against ElevenLabs, and the honest reason to route narration through a paid API at all is the same one the KB keeps landing on: local TTS - Chatterbox, Kokoro - is excellent in English and shakier across many languages, and production work doesn't always get to be in English. Turbo is the volume tier: same inputs, same voices, a little less polish than HD.

How it works

FAL_KEY from the .env beside your launcher; validate; estimate the cost from the character count and compare it to max_cost_usd; submit one billed request; poll with a live status on the node; download to output/fal/fal-ai_minimax_speech-2.6-turbo/. There are no uploads in a TTS run, so a rejected input is a non-event. Every request is logged, and results are saved where you expect.

Inputs

• prompt - the script. 1 to 10,000 characters, and it accepts pauses written as <#x#>, x being seconds from 0.01 to 99.99, placed between speakable segments and never twice in a row. This is the whole bill: 1,000 characters is roughly 180 English words.
• voice_setting_voice_id - free text, default Wise_Woman, from MiniMax's predefined voices. Empty leaves it to the model. Not a clone slot - that's a separate node family in this pack.
• voice_setting_emotion - (not set), happy, sad, angry, fearful, disgusted, surprised, neutral.
• language_boost - set it explicitly (the list runs to a few dozen languages plus auto). Mismatched language hints are the usual cause of a read that sounds like it's guessing.
• 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; turn it on for script with numbers, units and abbreviations).
• Loudness: normalization_setting_enabled (on by default), normalization_setting_target_loudness (−18 LUFS default, −70 to −10), normalization_setting_target_range (8 LU default), normalization_setting_target_peak (−0.5 dBTP 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 text replacements for the names it keeps mangling), output_format (url or hex; keep the default).

Optional: max_cost_usd (default $20; the run stops before sending if the estimate is above it, 0 = no cap), max_concurrent (default 1 - one paid fal call at a time in the run), reuse_identical_run (default on - identical inputs return the already-paid file, restart included), extra_json (raw JSON merged into fal's request last, for fields without a box).

Outputs: audio (a real waveform with a player on the node), audio_path, duration_ms (real read length - hand it to Narration Length to size shots, or Word Timings to place captions), 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, next to run_nvidia_gpu.bat:

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

Restart and you'll find it under arkennemasis/fal/Audio/MiniMax.

Notes

The normalization being on by default is the quiet win here. Twenty clips generated across an afternoon all land at the same loudness, which means your assembly doesn't need a level-matching pass per clip - and if you're feeding clips into the pack's Video Assemble, that's one fewer thing to fight.

Two habits: put the pauses in on the first pass rather than re-rendering to fix pacing, and don't expect the reuse logic to be invisible. Same text, same settings, and you get your earlier file back because the node refuses to charge you twice. Change one word and it's a fresh 6-cent generation.