Realtime
ModelsExpand Collapse
class ConversationItemAdded: …Sent by the server when an Item is added to the default Conversation. This can happen in several cases:
- When the client sends a
conversation.item.create event.
- When the input audio buffer is committed. In this case the item will be a user message containing the audio from the buffer.
- When the model is generating a Response. In this case the
conversation.item.added event will be sent when the model starts generating a specific Item, and thus it will not yet have any content (and status will be in_progress).
The event will include the full content of the Item (except when model is generating a Response) except for audio data, which can be retrieved separately with a conversation.item.retrieve event if necessary.
Sent by the server when an Item is added to the default Conversation. This can happen in several cases:
- When the client sends a
conversation.item.createevent. - When the input audio buffer is committed. In this case the item will be a user message containing the audio from the buffer.
- When the model is generating a Response. In this case the
conversation.item.addedevent will be sent when the model starts generating a specific Item, and thus it will not yet have any content (andstatuswill bein_progress).
The event will include the full content of the Item (except when model is generating a Response) except for audio data, which can be retrieved separately with a conversation.item.retrieve event if necessary.
class ConversationItemCreateEvent: …Add a new Item to the Conversation’s context, including messages, function
calls, and function call responses. This event can be used both to populate a
“history” of the conversation and to add new items mid-stream, but has the
current limitation that it cannot populate assistant audio messages.
If successful, the server will respond with a conversation.item.created
event, otherwise an error event will be sent.
Add a new Item to the Conversation’s context, including messages, function calls, and function call responses. This event can be used both to populate a “history” of the conversation and to add new items mid-stream, but has the current limitation that it cannot populate assistant audio messages.
If successful, the server will respond with a conversation.item.created
event, otherwise an error event will be sent.
class ConversationItemCreatedEvent: …Returned when a conversation item is created. There are several scenarios that produce this event:
- The server is generating a Response, which if successful will produce
either one or two Items, which will be of type
message
(role assistant) or type function_call.
- The input audio buffer has been committed, either by the client or the
server (in
server_vad mode). The server will take the content of the
input audio buffer and add it to a new user message Item.
- The client has sent a
conversation.item.create event to add a new Item
to the Conversation.
Returned when a conversation item is created. There are several scenarios that produce this event:
- The server is generating a Response, which if successful will produce
either one or two Items, which will be of type
message(roleassistant) or typefunction_call. - The input audio buffer has been committed, either by the client or the
server (in
server_vadmode). The server will take the content of the input audio buffer and add it to a new user message Item. - The client has sent a
conversation.item.createevent to add a new Item to the Conversation.
class ConversationItemInputAudioTranscriptionCompletedEvent: …This event is the output of audio transcription for user audio written to the
user audio buffer. Transcription begins when the input audio buffer is
committed by the client or server (when VAD is enabled). Transcription runs
asynchronously with Response creation, so this event may come before or after
the Response events.
Realtime API models accept audio natively, and thus input transcription is a
separate process run on a separate ASR (Automatic Speech Recognition) model.
The transcript may diverge somewhat from the model’s interpretation, and
should be treated as a rough guide.
This event is the output of audio transcription for user audio written to the user audio buffer. Transcription begins when the input audio buffer is committed by the client or server (when VAD is enabled). Transcription runs asynchronously with Response creation, so this event may come before or after the Response events.
Realtime API models accept audio natively, and thus input transcription is a separate process run on a separate ASR (Automatic Speech Recognition) model. The transcript may diverge somewhat from the model’s interpretation, and should be treated as a rough guide.
class ConversationItemRetrieveEvent: …Send this event when you want to retrieve the server’s representation of a specific item in the conversation history. This is useful, for example, to inspect user audio after noise cancellation and VAD.
The server will respond with a conversation.item.retrieved event,
unless the item does not exist in the conversation history, in which case the
server will respond with an error.
Send this event when you want to retrieve the server’s representation of a specific item in the conversation history. This is useful, for example, to inspect user audio after noise cancellation and VAD.
The server will respond with a conversation.item.retrieved event,
unless the item does not exist in the conversation history, in which case the
server will respond with an error.
class ConversationItemTruncateEvent: …Send this event to truncate a previous assistant message’s audio. The server
will produce audio faster than realtime, so this event is useful when the user
interrupts to truncate audio that has already been sent to the client but not
yet played. This will synchronize the server’s understanding of the audio with
the client’s playback.
Truncating audio will delete the server-side text transcript to ensure there
is not text in the context that hasn’t been heard by the user.
If successful, the server will respond with a conversation.item.truncated
event.
Send this event to truncate a previous assistant message’s audio. The server will produce audio faster than realtime, so this event is useful when the user interrupts to truncate audio that has already been sent to the client but not yet played. This will synchronize the server’s understanding of the audio with the client’s playback.
Truncating audio will delete the server-side text transcript to ensure there is not text in the context that hasn’t been heard by the user.
If successful, the server will respond with a conversation.item.truncated
event.
class ConversationItemTruncatedEvent: …Returned when an earlier assistant audio message item is truncated by the
client with a conversation.item.truncate event. This event is used to
synchronize the server’s understanding of the audio with the client’s playback.
This action will truncate the audio and remove the server-side text transcript
to ensure there is no text in the context that hasn’t been heard by the user.
Returned when an earlier assistant audio message item is truncated by the
client with a conversation.item.truncate event. This event is used to
synchronize the server’s understanding of the audio with the client’s playback.
This action will truncate the audio and remove the server-side text transcript to ensure there is no text in the context that hasn’t been heard by the user.
class InputAudioBufferAppendEvent: …Send this event to append audio bytes to the input audio buffer. The audio
buffer is temporary storage you can write to and later commit. A “commit” will create a new
user message item in the conversation history from the buffer content and clear the buffer.
Input audio transcription (if enabled) will be generated when the buffer is committed.
If VAD is enabled the audio buffer is used to detect speech and the server will decide
when to commit. When Server VAD is disabled, you must commit the audio buffer
manually. Input audio noise reduction operates on writes to the audio buffer.
The client may choose how much audio to place in each event up to a maximum
of 15 MiB, for example streaming smaller chunks from the client may allow the
VAD to be more responsive. Unlike most other client events, the server will
not send a confirmation response to this event.
Send this event to append audio bytes to the input audio buffer. The audio buffer is temporary storage you can write to and later commit. A “commit” will create a new user message item in the conversation history from the buffer content and clear the buffer. Input audio transcription (if enabled) will be generated when the buffer is committed.
If VAD is enabled the audio buffer is used to detect speech and the server will decide when to commit. When Server VAD is disabled, you must commit the audio buffer manually. Input audio noise reduction operates on writes to the audio buffer.
The client may choose how much audio to place in each event up to a maximum of 15 MiB, for example streaming smaller chunks from the client may allow the VAD to be more responsive. Unlike most other client events, the server will not send a confirmation response to this event.
class InputAudioBufferCommitEvent: …Send this event to commit the user input audio buffer, which will create a new user message item in the conversation. This event will produce an error if the input audio buffer is empty. When in Server VAD mode, the client does not need to send this event, the server will commit the audio buffer automatically.
Committing the input audio buffer will trigger input audio transcription (if enabled in session configuration), but it will not create a response from the model. The server will respond with an input_audio_buffer.committed event.
Send this event to commit the user input audio buffer, which will create a new user message item in the conversation. This event will produce an error if the input audio buffer is empty. When in Server VAD mode, the client does not need to send this event, the server will commit the audio buffer automatically.
Committing the input audio buffer will trigger input audio transcription (if enabled in session configuration), but it will not create a response from the model. The server will respond with an input_audio_buffer.committed event.
class InputAudioBufferDtmfEventReceivedEvent: …SIP Only: Returned when an DTMF event is received. A DTMF event is a message that
represents a telephone keypad press (0–9, *, #, A–D). The event property
is the keypad that the user press. The received_at is the UTC Unix Timestamp
that the server received the event.
SIP Only: Returned when an DTMF event is received. A DTMF event is a message that
represents a telephone keypad press (0–9, *, #, A–D). The event property
is the keypad that the user press. The received_at is the UTC Unix Timestamp
that the server received the event.
class InputAudioBufferSpeechStartedEvent: …Sent by the server when in server_vad mode to indicate that speech has been
detected in the audio buffer. This can happen any time audio is added to the
buffer (unless speech is already detected). The client may want to use this
event to interrupt audio playback or provide visual feedback to the user.
The client should expect to receive a input_audio_buffer.speech_stopped event
when speech stops. The item_id property is the ID of the user message item
that will be created when speech stops and will also be included in the
input_audio_buffer.speech_stopped event (unless the client manually commits
the audio buffer during VAD activation).
Sent by the server when in server_vad mode to indicate that speech has been
detected in the audio buffer. This can happen any time audio is added to the
buffer (unless speech is already detected). The client may want to use this
event to interrupt audio playback or provide visual feedback to the user.
The client should expect to receive a input_audio_buffer.speech_stopped event
when speech stops. The item_id property is the ID of the user message item
that will be created when speech stops and will also be included in the
input_audio_buffer.speech_stopped event (unless the client manually commits
the audio buffer during VAD activation).
class InputAudioBufferTimeoutTriggered: …Returned when the Server VAD timeout is triggered for the input audio buffer. This is configured
with idle_timeout_ms in the turn_detection settings of the session, and it indicates that
there hasn’t been any speech detected for the configured duration.
The audio_start_ms and audio_end_ms fields indicate the segment of audio after the last
model response up to the triggering time, as an offset from the beginning of audio written
to the input audio buffer. This means it demarcates the segment of audio that was silent and
the difference between the start and end values will roughly match the configured timeout.
The empty audio will be committed to the conversation as an input_audio item (there will be a
input_audio_buffer.committed event) and a model response will be generated. There may be speech
that didn’t trigger VAD but is still detected by the model, so the model may respond with
something relevant to the conversation or a prompt to continue speaking.
Returned when the Server VAD timeout is triggered for the input audio buffer. This is configured
with idle_timeout_ms in the turn_detection settings of the session, and it indicates that
there hasn’t been any speech detected for the configured duration.
The audio_start_ms and audio_end_ms fields indicate the segment of audio after the last
model response up to the triggering time, as an offset from the beginning of audio written
to the input audio buffer. This means it demarcates the segment of audio that was silent and
the difference between the start and end values will roughly match the configured timeout.
The empty audio will be committed to the conversation as an input_audio item (there will be a
input_audio_buffer.committed event) and a model response will be generated. There may be speech
that didn’t trigger VAD but is still detected by the model, so the model may respond with
something relevant to the conversation or a prompt to continue speaking.
class OutputAudioBufferClearEvent: …WebRTC/SIP Only: Emit to cut off the current audio response. This will trigger the server to
stop generating audio and emit a output_audio_buffer.cleared event. This
event should be preceded by a response.cancel client event to stop the
generation of the current response.
Learn more.
WebRTC/SIP Only: Emit to cut off the current audio response. This will trigger the server to
stop generating audio and emit a output_audio_buffer.cleared event. This
event should be preceded by a response.cancel client event to stop the
generation of the current response.
Learn more.
class RateLimitsUpdatedEvent: …Emitted at the beginning of a Response to indicate the updated rate limits.
When a Response is created some tokens will be “reserved” for the output
tokens, the rate limits shown here reflect that reservation, which is then
adjusted accordingly once the Response is completed.
Emitted at the beginning of a Response to indicate the updated rate limits. When a Response is created some tokens will be “reserved” for the output tokens, the rate limits shown here reflect that reservation, which is then adjusted accordingly once the Response is completed.
Optional[RealtimeAudioInputTurnDetection]Configuration for turn detection, ether Server VAD or Semantic VAD. This can be set to null to turn off, in which case the client must manually trigger model response.
Server VAD means that the model will detect the start and end of speech based on audio volume and respond at the end of user speech.
Semantic VAD is more advanced and uses a turn detection model (in conjunction with VAD) to semantically estimate whether the user has finished speaking, then dynamically sets a timeout based on this probability. For example, if user audio trails off with “uhhm”, the model will score a low probability of turn end and wait longer for the user to continue speaking. This can be useful for more natural conversations, but may have a higher latency.
For gpt-realtime-whisper transcription sessions, turn detection must be
set to null; VAD is not supported.
Configuration for turn detection, ether Server VAD or Semantic VAD. This can be set to null to turn off, in which case the client must manually trigger model response.
Server VAD means that the model will detect the start and end of speech based on audio volume and respond at the end of user speech.
Semantic VAD is more advanced and uses a turn detection model (in conjunction with VAD) to semantically estimate whether the user has finished speaking, then dynamically sets a timeout based on this probability. For example, if user audio trails off with “uhhm”, the model will score a low probability of turn end and wait longer for the user to continue speaking. This can be useful for more natural conversations, but may have a higher latency.
For gpt-realtime-whisper transcription sessions, turn detection must be
set to null; VAD is not supported.
class RealtimeConversationItemSystemMessage: …A system message in a Realtime conversation can be used to provide additional context or instructions to the model. This is similar but distinct from the instruction prompt provided at the start of a conversation, as system messages can be added at any point in the conversation. For major changes to the conversation’s behavior, use instructions, but for smaller updates (e.g. “the user is now asking about a different topic”), use system messages.
A system message in a Realtime conversation can be used to provide additional context or instructions to the model. This is similar but distinct from the instruction prompt provided at the start of a conversation, as system messages can be added at any point in the conversation. For major changes to the conversation’s behavior, use instructions, but for smaller updates (e.g. “the user is now asking about a different topic”), use system messages.
class RealtimeResponseCreateMcpTool: …Give the model access to additional tools via remote Model Context Protocol
(MCP) servers. Learn more about MCP.
Give the model access to additional tools via remote Model Context Protocol (MCP) servers. Learn more about MCP.
class RealtimeResponseUsageInputTokenDetails: …Details about the input tokens used in the Response. Cached tokens are tokens from previous turns in the conversation that are included as context for the current response. Cached tokens here are counted as a subset of input tokens, meaning input tokens will include cached and uncached tokens.
Details about the input tokens used in the Response. Cached tokens are tokens from previous turns in the conversation that are included as context for the current response. Cached tokens here are counted as a subset of input tokens, meaning input tokens will include cached and uncached tokens.
List[RealtimeToolsConfigUnion]Tools available to the model.
Tools available to the model.
Give the model access to additional tools via remote Model Context Protocol
(MCP) servers. Learn more about MCP.
Give the model access to additional tools via remote Model Context Protocol (MCP) servers. Learn more about MCP.
Optional[RealtimeTracingConfig]Realtime API can write session traces to the Traces Dashboard. Set to null to disable tracing. Once
tracing is enabled for a session, the configuration cannot be modified.
auto will create a trace for the session with default values for the
workflow name, group id, and metadata.
Realtime API can write session traces to the Traces Dashboard. Set to null to disable tracing. Once tracing is enabled for a session, the configuration cannot be modified.
auto will create a trace for the session with default values for the
workflow name, group id, and metadata.
Optional[RealtimeTranscriptionSessionAudioInputTurnDetection]Configuration for turn detection, ether Server VAD or Semantic VAD. This can be set to null to turn off, in which case the client must manually trigger model response.
Server VAD means that the model will detect the start and end of speech based on audio volume and respond at the end of user speech.
Semantic VAD is more advanced and uses a turn detection model (in conjunction with VAD) to semantically estimate whether the user has finished speaking, then dynamically sets a timeout based on this probability. For example, if user audio trails off with “uhhm”, the model will score a low probability of turn end and wait longer for the user to continue speaking. This can be useful for more natural conversations, but may have a higher latency.
For gpt-realtime-whisper transcription sessions, turn detection must be
set to null; VAD is not supported.
Configuration for turn detection, ether Server VAD or Semantic VAD. This can be set to null to turn off, in which case the client must manually trigger model response.
Server VAD means that the model will detect the start and end of speech based on audio volume and respond at the end of user speech.
Semantic VAD is more advanced and uses a turn detection model (in conjunction with VAD) to semantically estimate whether the user has finished speaking, then dynamically sets a timeout based on this probability. For example, if user audio trails off with “uhhm”, the model will score a low probability of turn end and wait longer for the user to continue speaking. This can be useful for more natural conversations, but may have a higher latency.
For gpt-realtime-whisper transcription sessions, turn detection must be
set to null; VAD is not supported.
class RealtimeTranslationInputAudioBufferAppendEvent: …Send this event to append audio bytes to the translation session input audio buffer.
WebSocket translation sessions accept base64-encoded 24 kHz PCM16 mono
little-endian raw audio bytes. Unsupported websocket audio formats return a
validation error because lower-quality audio materially degrades translation
quality.
Translation consumes 200 ms engine frames. For best realtime behavior, append
audio in 200 ms chunks. If a chunk is shorter, the server buffers it until it
has enough audio for one frame. If a chunk is longer, the server splits it into
200 ms frames and enqueues them back-to-back.
Keep appending silence while the session is active. If a client stops sending
audio and later resumes, model time treats the resumed audio as contiguous with
the previous audio rather than as a real-world pause.
Send this event to append audio bytes to the translation session input audio buffer.
WebSocket translation sessions accept base64-encoded 24 kHz PCM16 mono little-endian raw audio bytes. Unsupported websocket audio formats return a validation error because lower-quality audio materially degrades translation quality.
Translation consumes 200 ms engine frames. For best realtime behavior, append audio in 200 ms chunks. If a chunk is shorter, the server buffers it until it has enough audio for one frame. If a chunk is longer, the server splits it into 200 ms frames and enqueues them back-to-back.
Keep appending silence while the session is active. If a client stops sending audio and later resumes, model time treats the resumed audio as contiguous with the previous audio rather than as a real-world pause.
class RealtimeTranslationInputTranscriptDeltaEvent: …Returned when optional source-language transcript text is available. This event
is emitted only when audio.input.transcription is configured.
Transcript deltas are append-only text fragments. Clients should not insert
unconditional spaces between deltas.
Returned when optional source-language transcript text is available. This event
is emitted only when audio.input.transcription is configured.
Transcript deltas are append-only text fragments. Clients should not insert unconditional spaces between deltas.
When the number of tokens in a conversation exceeds the model’s input token limit, the conversation be truncated, meaning messages (starting from the oldest) will not be included in the model’s context. A 32k context model with 4,096 max output tokens can only include 28,224 tokens in the context before truncation occurs.
Clients can configure truncation behavior to truncate with a lower max token limit, which is an effective way to control token usage and cost.
Truncation will reduce the number of cached tokens on the next turn (busting the cache), since messages are dropped from the beginning of the context. However, clients can also configure truncation to retain messages up to a fraction of the maximum context size, which will reduce the need for future truncations and thus improve the cache rate.
Truncation can be disabled entirely, which means the server will never truncate but would instead return an error if the conversation exceeds the model’s input token limit.
When the number of tokens in a conversation exceeds the model’s input token limit, the conversation be truncated, meaning messages (starting from the oldest) will not be included in the model’s context. A 32k context model with 4,096 max output tokens can only include 28,224 tokens in the context before truncation occurs.
Clients can configure truncation behavior to truncate with a lower max token limit, which is an effective way to control token usage and cost.
Truncation will reduce the number of cached tokens on the next turn (busting the cache), since messages are dropped from the beginning of the context. However, clients can also configure truncation to retain messages up to a fraction of the maximum context size, which will reduce the need for future truncations and thus improve the cache rate.
Truncation can be disabled entirely, which means the server will never truncate but would instead return an error if the conversation exceeds the model’s input token limit.
class ResponseCancelEvent: …Send this event to cancel an in-progress response. The server will respond
with a response.done event with a status of response.status=cancelled. If
there is no response to cancel, the server will respond with an error. It’s safe
to call response.cancel even if no response is in progress, an error will be
returned the session will remain unaffected.
Send this event to cancel an in-progress response. The server will respond
with a response.done event with a status of response.status=cancelled. If
there is no response to cancel, the server will respond with an error. It’s safe
to call response.cancel even if no response is in progress, an error will be
returned the session will remain unaffected.
class ResponseCreateEvent: …This event instructs the server to create a Response, which means triggering
model inference. When in Server VAD mode, the server will create Responses
automatically.
A Response will include at least one Item, and may have two, in which case
the second will be a function call. These Items will be appended to the
conversation history by default.
The server will respond with a response.created event, events for Items
and content created, and finally a response.done event to indicate the
Response is complete.
The response.create event includes inference configuration like
instructions and tools. If these are set, they will override the Session’s
configuration for this Response only.
Responses can be created out-of-band of the default Conversation, meaning that they can
have arbitrary input, and it’s possible to disable writing the output to the Conversation.
Only one Response can write to the default Conversation at a time, but otherwise multiple
Responses can be created in parallel. The metadata field is a good way to disambiguate
multiple simultaneous Responses.
Clients can set conversation to none to create a Response that does not write to the default
Conversation. Arbitrary input can be provided with the input field, which is an array accepting
raw Items and references to existing Items.
This event instructs the server to create a Response, which means triggering model inference. When in Server VAD mode, the server will create Responses automatically.
A Response will include at least one Item, and may have two, in which case the second will be a function call. These Items will be appended to the conversation history by default.
The server will respond with a response.created event, events for Items
and content created, and finally a response.done event to indicate the
Response is complete.
The response.create event includes inference configuration like
instructions and tools. If these are set, they will override the Session’s
configuration for this Response only.
Responses can be created out-of-band of the default Conversation, meaning that they can
have arbitrary input, and it’s possible to disable writing the output to the Conversation.
Only one Response can write to the default Conversation at a time, but otherwise multiple
Responses can be created in parallel. The metadata field is a good way to disambiguate
multiple simultaneous Responses.
Clients can set conversation to none to create a Response that does not write to the default
Conversation. Arbitrary input can be provided with the input field, which is an array accepting
raw Items and references to existing Items.
class ResponseDoneEvent: …Returned when a Response is done streaming. Always emitted, no matter the
final state. The Response object included in the response.done event will
include all output Items in the Response but will omit the raw audio data.
Clients should check the status field of the Response to determine if it was successful
(completed) or if there was another outcome: cancelled, failed, or incomplete.
A response will contain all output items that were generated during the response, excluding
any audio content.
Returned when a Response is done streaming. Always emitted, no matter the
final state. The Response object included in the response.done event will
include all output Items in the Response but will omit the raw audio data.
Clients should check the status field of the Response to determine if it was successful
(completed) or if there was another outcome: cancelled, failed, or incomplete.
A response will contain all output items that were generated during the response, excluding any audio content.
class SessionUpdateEvent: …Send this event to update the session’s configuration.
The client may send this event at any time to update any field
except for voice and model. voice can be updated only if there have been no other audio outputs yet.
When the server receives a session.update, it will respond
with a session.updated event showing the full, effective configuration.
Only the fields that are present in the session.update are updated. To clear a field like
instructions, pass an empty string. To clear a field like tools, pass an empty array.
To clear a field like turn_detection, pass null.
Send this event to update the session’s configuration.
The client may send this event at any time to update any field
except for voice and model. voice can be updated only if there have been no other audio outputs yet.
When the server receives a session.update, it will respond
with a session.updated event showing the full, effective configuration.
Only the fields that are present in the session.update are updated. To clear a field like
instructions, pass an empty string. To clear a field like tools, pass an empty array.
To clear a field like turn_detection, pass null.
RealtimeCalls
Accept call
Create call
Refer call
Reject call
RealtimeClient Secrets
Create client secret
ModelsExpand Collapse
class RealtimeTranscriptionSessionTurnDetection: …Configuration for turn detection. Can be set to null to turn off. Server
VAD means that the model will detect the start and end of speech based on
audio volume and respond at the end of user speech. For gpt-realtime-whisper, this must be null; VAD is not supported.
Configuration for turn detection. Can be set to null to turn off. Server
VAD means that the model will detect the start and end of speech based on
audio volume and respond at the end of user speech. For gpt-realtime-whisper, this must be null; VAD is not supported.