arkennemasis fal · MiniMax Preview Speech 2.5 Turbo · $0.06/1k chars

The cheapest way to get speech out of the 2.5 preview family: $0.06 per 1,000 characters. Same voice list, same language picker, same inputs as the HD node one rung up. If you're producing narration in bulk - a series, a course, a pile of shorts - this is the one that keeps the budget boring.

What it is

A wrapper around fal's minimax/preview/speech-2.5-turbo endpoint, one node in the pack's 36-model fal audio catalogue. Turbo sits inside MiniMax's voice family, the line they benchmark against ElevenLabs - which is the whole reason to reach for a hosted TTS instead of hanging Chatterbox off your own GPU. Local TTS does generic English narration extremely well now; it gets shakier across many languages, and this node's language_boost picker is unusually broad for a cheap tier, including Persian, Filipino and Tamil alongside the usual Chinese/English/Japanese/Korean/European set.

How it works

FAL_KEY comes from the .env beside your ComfyUI launcher. The node validates the text, estimates the cost from its character count, submits one billed request to fal, polls with a live status on the node, and downloads into output/fal/fal-ai_minimax_preview_speech-2.5-turbo/. Nothing is uploaded on a TTS run, so every failure before the submit is free.

Inputs

• text - named text, not prompt; max 5000 characters. Cost maths: about 180 words per 1,000 characters, so a two-minute script is roughly 2,000 characters - a hair over 12 cents.
• voice_setting_voice_id - free-text box, default Wise_Woman. Paste a MiniMax predefined voice ID; leaving it empty hands the choice to the model.
• voice_setting_emotion - (not set) or one of happy, sad, angry, fearful, disgusted, surprised, neutral. (not set) is usually right.
• language_boost - pick the language. Leaving it at (not set) with non-English text is the single most common cause of a "bad" result.
• voice_setting_speed - 0.5 to 2.0. With voice_setting_pitch (−12 to 12), voice_setting_vol (0.01–10) and voice_setting_english_normalization (off by default; on for number-heavy scripts).
• Output settings - 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), output_format (url or hex, non-streaming only; leave the default alone).

Optional: max_cost_usd (default $20 - the node stops before sending anything if the estimate is above it, 0 turns the cap off; at 6 cents per 1,000 characters you'd need a very long script to hit it), max_concurrent (default 1, one paid fal call at a time across a run), reuse_identical_run (default on), extra_json.

Outputs: audio (waveform + sample rate, playable on the node), audio_path, duration_ms (real measured length in milliseconds - feed it to Narration Length or Word Timings so cuts and captions follow the voice), and info.

Install

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

Or Install via Git URL in ComfyUI Manager. Then beside run_nvidia_gpu.bat:

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

Restart - it's under arkennemasis/fal/Audio/MiniMax, titled with its rate.

Practical notes

There are no pause markers on this generation. The <#x#> syntax arrives with the 2.6 and 2.8 nodes; here you get punctuation and line breaks, nothing else. If your read needs a dramatic beat, write two sentences and cut the silence in afterwards.

Two pack behaviours worth knowing before you conclude something's broken. reuse_identical_run returns the existing file when text and settings are unchanged - free, and correct, but it looks like nothing happened. And if you're rendering a batch of voice-overs through several fal nodes at once, leave max_concurrent at 1 unless you enjoy rate-limit errors; the default exists because ComfyUI runs nodes side by side.