Thanks — I’m happy to help with this. You can play a pre-recorded audio clip automatically when a call is answered, then let the agent continue the conversation. Below I’ll cover how to do it, whether WebM is supported, and practical implementation steps. Acknowledgement • You asked: 1) Can a pre-recorded audio play when the user picks up? 2) Is WebM supported? 3) How do you play the audio? Short answer • Yes — play the pre-recorded audio on answer using the agent/session API (use

session.say()

with an audio generator that yields

rtc.AudioFrame

). • WebM is supported only when the audio inside the WebM is Opus-encoded (Opus is the codec WebRTC/LiveKit expects). Converting to WAV can simplify decoding and playback. How to implement (high level) 1. Store the file (local disk or cloud). If remote, download it on the server before playback. 2. Decode the audio file into raw frames and create an

rtc.AudioFrame

(or equivalent structure used by your SDK). You can decode WebM/Opus with ffmpeg/PyAV, or load WAV directly with

wave

for ease. 3. Use

session.say()

and pass an async generator that yields the

rtc.AudioFrame

. Set

text=None

so only the audio plays. 4. After playback, let the agent continue (call

generate_reply()

or otherwise start the agent’s normal speech-to-text/response flow). Practical notes and code-patterns • Playing audio only (no TTS): ◦ Use

session.say(text=None, audio=audio_generator())

, where

audio_generator()

yields

rtc.AudioFrame

frames. • WebM support: ◦ WebM is supported if the contained audio is Opus. If your WebM uses Opus, decode it server-side to produce frames for

rtc.AudioFrame

. If you prefer simpler handling, convert to WAV and load the WAV frames directly. • Recommended approach: ◦ For reliability and simpler decoding on the backend, convert files to WAV (if feasible). WAV is straightforward to load and convert into audio frames. • Example patterns (pseudo/Python): a. Decode file into frames (WAV example shown conceptually):

with wave.open(str(audio_path), 'rb') as wav_file:
         frames = wav_file.readframes(wav_file.getnframes())
         sample_rate = wav_file.getframerate()
         num_channels = wav_file.getnchannels()

     audio_frame = rtc.AudioFrame(
         data=frames,
         sample_rate=sample_rate,
         num_channels=num_channels,
         samples_per_channel=wav_file.getnframes()
     )

1. Yield frames in an async generator and call `session.say()`:

async def audio_generator():
         yield audio_frame

     await session.say(text=None, audio=audio_generator())

1. After playback, start the agent conversation:

await session.generate_reply()

Questions to narrow your setup • Which SDK and language are you using (e.g., Python, Node, JS)? I can give a concrete code example for that environment. • Is your pre-recorded file currently WebM/Opus, WAV, or MP3? If WebM, do you know whether it’s Opus-encoded? • Where do you store the audio (local server, S3, other)? That affects whether you need to download/stream and how you decode. If you tell me your SDK/language and file format/storage, I’ll provide a short, copy-paste example that fits your stack and sample code to decode WebM (Opus) or play WAV. Sources: Playing Audio | LiveKit Documentation | Universal Egress | Agent speech and audio | LiveKit Documentation