OpenAI Read Aloud

Read a short text aloud with an editable voice and delivery prompt. The graph passes the supplied wording directly to speech synthesis; it does not summarize or rewrite it. The voice is AI-generated.

Try it

1. Import the graph and add the graph-scoped secret OPENAI_API_KEY in Graph Settings > Execution Profile.
2. Edit Text to read, Voice direction, and optionally Voice. The default voice is marin; cedar is another option.
3. Run. Bind Audio Preview to Generate speech → media. The transcript graph output contains the supplied script, and response serves the MP3.

The built-in sample is original fictional text. No upload is required, and this example works on a localhost dev instance with outbound access to OpenAI.

Graph and limits

Text + voice + delivery prompt → validate → speech body → HTTP request → audio

• One paid POST https://api.openai.com/v1/audio/speech call using gpt-4o-mini-tts. Script and voice instructions are sent to OpenAI.
• Script: 1–3500 characters, including at least one non-whitespace character. Voice instructions: 1–1000 characters. Voice must be one of the built-in options listed in the validation node. These are example limits, not provider maxima.
• Timeout: 120 seconds. Response cap: 10,000,000 bytes. One audio track; no loops, automatic retries, chapter splitting, concatenation, or background jobs.
• A validation failure stops before the paid call. Each regex check feeds its match collection into JSON Path 0; an empty collection deliberately fails. The failing node’s title identifies the limit to fix.
• Inspect Generate speech → status/json for provider errors such as missing credentials, model access, or rate limits. A failed request produces no audio.
• Audio uses the runtime’s existing media/artifact behavior. No generated media or credentials are included in this export. Running again can incur another charge. There is no publication, messaging, or deletion step.

To read text from a PDF without adaptation, paste a short extracted passage into Text to read. For PDF upload and prompted adaptation, use OpenAI PDF Narration.

Validation

The graph is checked against the sibling runtime with mocked provider responses. Live voice quality and paid API execution still need testing in your instance.

OpenAI speech guide

To repeat offline checks from the examples repository:

python3 scripts/check_narration_examples.py --runtime-root ../subfork-new/backend/runtime

The runtime path is a CLI argument and can point to another checkout. No provider credentials or network access are used by these checks.