Inject Audio into Live Calls on Asterisk and FreeSWITCH

Why inject audio into a live call

Most call audio comes from the people on the call. But many real systems need to add something while the call is running, without disconnecting anyone or moving them to a new context.

First decide who should hear it

A two-party call has two legs: the caller (often called the A leg) and the agent or callee (the B leg). Audio sent "to a leg" is heard by the person on that leg. That gives three common targets:

A second choice matters too: should the injected audio replace the conversation while it plays, or mix with it? A recording notice usually replaces; a beep or a whisper usually mixes, so the conversation continues underneath.

Asterisk: the main techniques

Plain dialplan applications such as Playback work before a call is bridged. Once two parties are bridged, live injection is easiest through ARI (the Asterisk REST Interface), which can act on channels and bridges at any time.

Asterisk examples

Play a notice to everyone in a bridge with ARI. Your ARI application knows the bridge ID; one request plays the file to all participants.

# Play a recording notice to everyone in a bridge
curl -u ariuser:secret -X POST \
  "http://127.0.0.1:8088/ari/bridges/$BRIDGE_ID/play?media=sound:custom/recording-notice"

# Play only to the caller's channel
curl -u ariuser:secret -X POST \
  "http://127.0.0.1:8088/ari/channels/$CALLER_CHANNEL/play?media=sound:custom/transfer-notice"

Whisper to the agent with an ARI snoop channel. A snoop channel attached to the agent's channel with whisper=out (audio going out from Asterisk to the agent) sends whatever it plays to the agent only. Whispering needs a continuous media stream to the channel, which a bridged call normally provides. Play a file, a text-to-speech prompt, or bridge a supervisor's channel into it.

# Create a snoop channel on the agent's channel: whisper to the agent, don't listen
curl -u ariuser:secret -X POST \
  "http://127.0.0.1:8088/ari/channels/$AGENT_CHANNEL/snoop?app=coach&whisper=out&spy=none"
# -> returns the snoop channel; then play a hint into it:
curl -u ariuser:secret -X POST \
  "http://127.0.0.1:8088/ari/channels/$SNOOP_CHANNEL/play?media=sound:custom/offer-discount-hint"

Supervisor whisper and barge with ChanSpy. A supervisor dials an extension, enters the agent's number, and listens; the options decide whether they can talk to the agent only or to both.

; extensions.conf: supervisor features
exten => 555,1,NoOp(Whisper to agent)
 same => n,Read(AGENT,agent-number,4)
 same => n,ChanSpy(PJSIP/${AGENT},qw)   ; q = quiet, w = whisper to the agent only

exten => 556,1,NoOp(Barge into the call)
 same => n,Read(AGENT,agent-number,4)
 same => n,ChanSpy(PJSIP/${AGENT},qB)   ; B = barge: talk to both parties

FreeSWITCH: the main techniques

All of these can be triggered from the Event Socket (ESL), so a Node.js or Python service can inject audio into any active call by its UUID. Our FreeSWITCH ESL with Node.js guide covers the control side.

FreeSWITCH examples

# From fs_cli or an ESL "api" command, using the call's UUID

# Play a notice to both legs (interrupts the conversation while it plays)
uuid_broadcast 7d1c...e2 /var/sounds/recording-notice.wav both

# Play only to the caller (A leg)
uuid_broadcast 7d1c...e2 /var/sounds/please-hold.wav aleg

# Mix a short beep into the call without interrupting it
uuid_displace 7d1c...e2 start /var/sounds/beep.wav 0 mux

# Stop a running displace
uuid_displace 7d1c...e2 stop /var/sounds/beep.wav

For supervisors, eavesdrop lets a supervisor listen to a call and switch between listening, whispering to one side and talking to both using keypad digits during the session, which is the FreeSWITCH equivalent of ChanSpy.

Files, text-to-speech or live audio?

What you inject matters as much as how. Each source has a different cost in delay, flexibility and quality.

Most systems use a mix: recordings for anything legally worded, cached text-to-speech for common prompts, and live generation only where the content changes per call.

Injecting AI speech into calls

An AI voice agent speaking into a call is a special case: the audio is generated on the fly, sentence by sentence, and must stop the moment the caller starts talking.

• Stream call audio to the AI over AudioSocket or ARI external media on Asterisk, or mod_audio_stream on FreeSWITCH.
• Send the AI's speech back in small chunks (typically 20 ms frames) as soon as text-to-speech produces them.
• Match the call's audio format: for example 8 kHz signed linear for AudioSocket, resampling text-to-speech output if needed.
• On barge-in, stop sending audio immediately and clear any buffered audio, so the agent doesn't talk over the caller.
• Alternatively, add the AI as a participant in a conference or an SFU room such as LiveKit, connected over SIP.

Our guides to Asterisk AudioSocket and FreeSWITCH with OpenAI Realtime walk through both platforms in detail.

Comparing the methods

Getting the audio quality right

Building audio injection as a service

In production, injection is rarely typed into a console. It is triggered by events: a supervisor clicks "whisper", a timer fires a recording beep, a CRM rule decides an agent needs a hint. That calls for a small control service between your applications and the telephony platform.

Keeping the platform details behind one API also makes it easier to support both Asterisk and FreeSWITCH, or to move between them later.

Recording and monitoring notices, done right

• Play the notice to both parties as soon as the call is answered, before the conversation starts.
• Use the caller's language, chosen from the number they called or their account.
• Keep it short and clear; long legal wording makes callers hang up.
• If recording starts mid-call, for example on transfer, play the notice again at that point.
• For periodic beeps, mix them in at a low level rather than interrupting the conversation.
• Check the rules where your callers are: some places require notices or consent from everyone on the call.

Common pitfalls

• Playing to the wrong leg: confirm which leg is which in transfers and forwarded calls, where they can swap.
• Breaking the bridge: some dialplan approaches unbridge the call; use ARI or ESL commands that keep the call connected.
• Long announcements during urgent calls: keep notices short, and skip non-essential ones on repeat calls.
• Whispers leaking: test that 'agent only' audio really isn't heard by the caller, including on conference calls.
• Forgotten legal notices: recording and monitoring notices are required in many places; check where your callers are.
• Race conditions: injecting before the call is answered or after it ends causes errors; check the call state first.

Testing checklist

A real example

In CelloIP's AI voice agent for a contact centre, calls run through the centre's FreeSWITCH platform and the AI agent joins over SIP. When the agent hands a call to a person, the human agent receives a summary of the conversation so the caller doesn't repeat themselves, the same principle as an agent-only whisper: information for one party, without interrupting the other.

Frequently asked questions