zhenzhen-seed-speech-tts-v2-fal

ByteDance's Seed Speech TTS has become a favorite for "give me a clean, natural voice without ElevenLabs prices," and this node is the pack's door to its V2 model via the fal queue. Comfly_seed_speech_tts_v2_fal takes your text and returns a real ComfyUI AUDIO tensor you can drop straight into a video track or a save node - with 21 built-in voices, sample-rate control, and a voice_instruction field for nudging delivery. If you've been generating silent videos and want narration without renting a GPU or signing up for yet another TTS subscription, this is the node.

Like every FAL node here, the token goes in the node's own api_key field and traffic runs through zhenzhen's proxy (https://ai.t8star.org/fal in front of queue.fal.run). Submit to the fal-ai/bytedance/seed-speech/tts/v2 endpoint, poll, download. It's pay-per-call and prepaid-credit style, like all the fal routes.

The inputs that matter

• text - required, the thing to speak. Default is a test sentence.
• voice - 21 choices baked into the node, from stokie_en (default) to monkey_king_zh. The names are honest about their languages - bonnie_zh, sophie_en_zh, vienna_mixed_en_zh are all in there.
• output_format - mp3 or opus.
• sample_rate - 8000 to 48000, 24000 default. 48k if you're mastering; 24k is fine for web.
• speed (0.5–2.0), volume, pitch (-12 to +12 semitones) - the usual trim controls.
• language - auto (default) or zh/en/ja/es-mx/id/pt-br/ko/it/de/fr. Auto is usually right, but locking it can help with proper names.
• voice_instruction - the hidden gem: a free-text style instruction like "speak softly and slowly, like a bedtime story." Left empty when not set.

Outputs: audio (the tensor), audio_url, and response. Standard polling knobs (poll_interval 6s, max_poll_attempts 600 ≈ 1h) and the usual skip_error escape hatch.

Installing it

cd ComfyUI/custom_nodes
git clone https://github.com/T8mars/Comfyui-zhenzhen

or ComfyUI Manager (search "Comfyui-zhenzhen"), restart. Paste a zhenzhen token (overseas https://ai.t8star.org) into api_key. No models, no VRAM - the pack's audio dependencies are minimal and Manager handles them.

Common issues

• Silent or empty audio output - skip_error is probably on and swallowing a failure. Off while debugging.
• Voice sounds off for the language - pick a voice that matches your language; forcing auto with a clearly bilingual voice can give odd prosody.
• TTS is slow to come back - the queue can back up; that's what the polling knobs are for, not a bug in the node.

It's not cloning, and it's not trying to be - this is "good natural TTS voices, metered by the character," the kind of thing the audio-generation side of the ecosystem keeps going back to the cloud for. Cheap, easy, and wired into your graph.