Chirp 3: HD Voices

Chirp is the text-to-speech node in the Vertex AI pack, and honestly it's the sleeper hit. Everyone installs this pack for Veo or Imagen and then discovers you get Google's Chirp 3 HD voices for free along the way. If you've ever stared at a finished video and realized it needs a voiceover, this is the node that keeps you in the graph instead of exporting and hitting elevenlabs.

What it does

You type text, pick a language and a voice, and the node calls Google's Cloud Text-to-Speech API and returns a proper ComfyUI AUDIO output. There are 34 languages in the dropdown (English US/UK/AU, all the major European and Asian ones) and 30 voices - named after stars and moons, because Google never met a naming scheme it didn't like. Achernar, Alnilam, Zephyr, Puck. The voices are the "Chirp 3 HD" line, which is Google's good tier, and it shows: natural prosody, no obvious robot edge, believable multilingual output.

How it works

The node builds a full voice identifier from your two dropdowns - pick English (United States) and Achernar and it calls the API with en-US-Chirp3-HD-Achernar. It requests uncompressed LINEAR16 audio, saves it to a temp WAV, loads it with torchaudio, and hands you a {waveform, sample_rate} dict. That's the standard ComfyUI AUDIO format, so it plugs straight into anything that accepts audio. The API client is created lazily on first run and cached on the node.

Inputs that matter

• project_id and location - your Google Cloud project and region. Defaults come from GOOGLE_CLOUD_PROJECT and GOOGLE_CLOUD_LOCATION (falling back to us-central1), so if you've done the pack's setup you usually don't touch these.
• language - 34 choices, from Arabic to Vietnamese.
• voice_name - the 30 Chirp 3 HD voices. There's no per-voice preview, so budget one run to audition a few.
• text - multiline, whatever you want spoken.

The output is one AUDIO tensor. Wire it into SaveAudio or the VHS audio playback nodes to actually hear it - this pack ships a video preview node but not an audio one, which is the one gap worth knowing about.

Installing

This is a pack node, so it comes with the whole Vertex AI setup, not just TTS. Install via ComfyUI Manager (search "Vertex AI"), or:

cd ComfyUI/custom_nodes
git clone https://github.com/khanhlvg/vertex-ai-comfyui-nodes.git
cd vertex-ai-comfyui-nodes
pip install -r requirements.txt

Then authenticate and set your project:

gcloud auth application-default login
export GOOGLE_CLOUD_PROJECT=your-project-id
export GOOGLE_CLOUD_LOCATION=us-central1

The requirements pull in google-cloud-texttospeech plus the other Google SDKs - no local model downloads, no VRAM, just a project with billing and the Cloud Text-to-Speech API enabled.

Where people get burned

• Nothing plays. The AUDIO output needs a downstream audio node. Add SaveAudio or VHS and the mystery disappears.
• API errors on first run. If you never ran gcloud auth application-default login, every call fails with an auth error before it touches a voice. This is the single most common failure across this whole pack.
• Cost. TTS is metered per character. A short VO is pocket change, but generating a full script at scale adds up - keep an eye on it.

One aside: because every node in this pack shares the same auth setup, if Chirp works for you, the Veo and Imagen nodes next to it will too.