October 9, 2026

Murf added word-level timestamps to its WebSocket streaming API for Falcon 2.

Set add_word_timestamps to true at the top level of the frame, next to text. Murf says to keep it out of voice_config, where it's ignored without an error. Once it's on, it applies to the whole connection. Falcon then sends each spoken word with start_s and end_s in seconds. These arrive as separate JSON frames mixed in with the audio frames, not inside the base64 audio field.

This helps most with Murf AI voice agents that handle barge-in. When a user interrupts, stop playback and send a clear frame with that turn's context_id. Then drop any audio and timestamps past the current playback position. The timestamps you keep are the words the user actually heard.

The word field is normalized spoken text, so "$50" comes back as "fifty dollars" and "2pm" as "two p m". Word timestamps work in every locale except Chinese, Japanese, and Korean.