MiniMax Speech 2.8 HD (Replicate)

MiniMax Speech 2.8 HD is the most controllable text-to-speech in this pack, and for narration work that's the whole game. Where most TTS nodes give you a voice dropdown and a speed slider, this one hands you an emotion selector (happy, sad, angry, calm, excited - ten of them), a pitch knob in semitones, pause markers you type straight into the text, and a voice list long enough to scroll. It also plays nicely with ComfyUI's native AUDIO type, which is the quiet technical detail that makes everything downstream work.

TTS is the "cheap part" of any audio pipeline - the expensive compute is the video the voice drives. This node fits that: it's a cloud TTS with real production polish (MiniMax's speech models are genuinely good), and because it returns AUDIO rather than a file path, you can wire it straight into a lipsync node or SaveAudio. The pack's README specifically calls out that pattern: TTS → AUDIO → lipsync nodes via upload.

How it works

Standard engine around minimax/speech-2.8-hd. Your text goes up, the synthesized audio comes back, and the node converts it into a ComfyUI AUDIO dict (waveform + sample_rate). The single output is named audio and is native ComfyUI AUDIO - not a path string. That's why it can feed SaveAudio directly or upload into a lipsync model without any file juggling.

The inputs that matter

• text (required) - up to 10,000 characters. The neat trick: type markers like <#0.5#> to insert a 0.5-second pause. This is how you get natural pacing without post-editing.
• voice_id - default English_Wiselady. Any MiniMax system voice, or a voice_id from MiniMax's voice-cloning model on Replicate. That's your cloning door: clone a voice elsewhere, paste its id here.
• emotion - auto by default, plus happy/sad/angry/fearful/disgusted/surprised/calm/fluent/neutral. This is the differentiator; a "calm" vs "happy" render of the same line is dramatically different.
• speed - 0.5–2.0x.
• pitch - ±12 semitones. Nudge a voice up or down without changing who it is.
• volume - 0–10, default 1.0 (the model's native gain).
• audio_format - mp3 (default), wav/flac for lossless, or pcm for raw bytes.
• channel - mono or stereo.
• sample_rate - up to 44100.
• bitrate - only applies to mp3.
• english_normalization - improves number/date reading for English, at a small latency cost.
• subtitle_enable - returns sentence timestamps (non-streaming only). Turn this on if you're syncing subtitles to the audio.
• language_boost - an optional language hint; None/Automatic or pick a locale.

Output: audio, native ComfyUI AUDIO.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/Pharma-Lobby/ComfyUI-Replicate-Select

Or ComfyUI Manager → Install Custom Nodes → "Replicate Select", restart. Dependency: replicate>=1.0.7. Auth via environment variable:

export REPLICATE_API_TOKEN=r8_...

Where people get burned

• Forgetting you can control delivery, not just content. The emotion dial and pause markers are the features; if your narration sounds flat it's because you're treating this like a basic TTS.
• Voice cloning needs its own model. The pack's nodes only use cloned voice ids - you clone through MiniMax's separate voice-cloning model, then paste the id here.
• mp3 vs pcm. The pcm format is raw bytes, useful for some pipelines and useless for others. Stick to wav for anything going into a lipsync node.