arkennemasis fal · MiniMax Speech 2.8 HD · $0.1/1k chars

Newest of the MiniMax speech nodes in this pack, and the first one with a personality dial. Same $0.10 per 1,000 characters as the rest of the HD tier, and it keeps everything 2.6 added - pause markers, loudness normalization, a 10,000-character script budget - then adds non-verbal interjections and a three-axis voice tweaker on top.

The two things that are new

Interjections. You can type (laughs), (sighs), (coughs), (clears throat), (gasps), (sniffs), (groans) or (yawns) into the script and the model performs them. For a character voice in a short - the sort of thing where a dry read kills the joke - that's a genuinely useful tool, and it's the kind of control you'd otherwise be farming out to an audio editor and a library of stingers.

voice_modify_pitch, voice_modify_intensity, voice_modify_timbre - three integers, each −100 to 100, defaulting to 0. These are not the same as voice_setting_pitch, which shifts the read by semitones; this trio re-timbres the voice itself: pitch character, energy, tonal quality. It's how you take one predefined voice and get three distinguishable speakers out of it.

How it works

The pack's usual fal plumbing: FAL_KEY read from the .env beside your launcher, input validation, a cost estimate from the character count checked against max_cost_usd, one billed submit, polling with a status label on the node, download into output/fal/fal-ai_minimax_speech-2.8-hd/. No uploads happen on a TTS run, so nothing before the submit can cost you anything. MiniMax's voice line is closed and hosted, benchmarked by the lab against ElevenLabs - the reason to pay for it rather than run Chatterbox locally is multilingual reliability, and that tradeoff hasn't changed.

Inputs

• prompt - script, 1 to 10,000 characters. Accepts <#x#> pauses (seconds, 0.01–99.99, between speakable segments, never consecutive) and the interjection tags above.
• voice_setting_voice_id - free-text, default Wise_Woman, from MiniMax's predefined IDs. Empty leaves it to the model.
• voice_setting_emotion - (not set), happy, sad, angry, fearful, disgusted, surprised, neutral.
• language_boost - the explicit language picker including auto. Set it.
• voice_setting_speed (0.5–2.0), voice_setting_pitch (−12 to 12 semitones - different dial from the voice_modify_* trio), voice_setting_vol (0.01–10), voice_setting_english_normalization (off by default).
• voice_modify_pitch / voice_modify_intensity / voice_modify_timbre - each −100 to 100, default 0.
• Loudness: normalization_setting_enabled (on by default), normalization_setting_target_loudness (−18 LUFS default), normalization_setting_target_range (8 LU), normalization_setting_target_peak (−0.5 dBTP).
• Output: audio_setting_format (mp3, pcm, flac), audio_setting_bitrate, audio_setting_sample_rate, audio_setting_channel (1 mono, 2 stereo), pronunciation_dict (JSON replacements), output_format (url/hex; leave the default).

Optional: max_cost_usd ($20 default, 0 = off; the node stops before sending anything if the estimate is above it), max_concurrent (default 1 - one paid fal call at a time across a run), reuse_identical_run (identical inputs return the file you already paid for, restart included), extra_json.

Outputs: audio (waveform + sample rate, with a player on the node), audio_path, duration_ms (feed it to Narration Length or Word Timings to time shots and captions to the real read), info (request id, cost 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; it's under arkennemasis/fal/Audio/MiniMax. There's also a Speech 2.8 Turbo sibling in the pack if you'd rather iterate at the Turbo rate.

Using it without wasting takes

Interjections are cheap fun and expensive habits - every one of them is another billed run while you fine-tune placement. Write them on the first pass, then re-render once, not seven times.

Where the voice_modify_* dials earn their keep: intensity is the one I'd touch first. Nudging it positive makes a read sound like the speaker cares, and it's the least likely of the three to break the voice's identity. Leave timbre alone unless you're deliberately building a second character from the same predefined voice.

And the standing pack behaviour: same text plus same settings returns your earlier file rather than billing again. If a re-render looks suspiciously identical, that's the node being honest, not lazy.