POST /realtime/sessions

Default value: "application/json"

Sampling temperature for the model, limited to [0.6, 1.2]. For audio models a temperature of 0.8 is highly recommended for best performance.

Default value: 0.8

Tools (functions) available to the model.

The name of the function.

The description of the function, including guidance on when and how to call it, and guidance about what to tell the user when calling (if anything).

Parameters of the function in JSON Schema.

The type of the tool, i.e. function.

Possible values:

• "function"

Maximum number of output tokens for a single assistant response, inclusive of tool calls. Provide an integer between 1 and 4096 to limit output tokens, or inf for the maximum available tokens for a given model. Defaults to inf.

How the model chooses tools. Options are auto, none, required, or specify a function.

Default value: "auto"

The Realtime model used for this session.

Possible values:

• "gpt-4o-mini-realtime-preview-2024-12-17"
• "gpt-4o-realtime-preview-2024-10-01"
• "gpt-4o-realtime-preview"
• "gpt-4o-mini-realtime-preview"
• "gpt-4o-realtime-preview-2024-12-17"

Configuration for input audio noise reduction. This can be set to null to turn off. Noise reduction filters audio added to the input audio buffer before it is sent to VAD and the model. Filtering the audio can improve VAD and turn detection accuracy (reducing false positives) and model performance by improving perception of the input audio.

Type of noise reduction. near_field is for close-talking microphones such as headphones, far_field is for far-field microphones such as laptop or conference room microphones.

Possible values:

• "near_field"
• "far_field"

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 conjuction 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.

Used only for server_vad mode. Amount of audio to include before the VAD detected speech (in milliseconds). Defaults to 300ms.

Used only for server_vad mode. Activation threshold for VAD (0.0 to 1.0), this defaults to 0.5. A higher threshold will require louder audio to activate the model, and thus might perform better in noisy environments.

Used only for semantic_vad mode. The eagerness of the model to respond. low will wait longer for the user to continue speaking, high will respond more quickly. auto is the default and is equivalent to medium.

Possible values:

• "medium"
• "low"
• "high"
• "auto"

Default value: "auto"

Type of turn detection.

Possible values:

• "server_vad"
• "semantic_vad"

Default value: "server_vad"

Used only for server_vad mode. Duration of silence to detect speech stop (in milliseconds). Defaults to 500ms. With shorter values the model will respond more quickly, but may jump in on short pauses from the user.

Whether or not to automatically interrupt any ongoing response with output to the default conversation (i.e. conversation of auto) when a VAD start event occurs.

Default value: true

Whether or not to automatically generate a response when a VAD stop event occurs.

Default value: true

The format of output audio. Options are pcm16, g711_ulaw, or g711_alaw. For pcm16, output audio is sampled at a rate of 24kHz.

Possible values:

• "g711_ulaw"
• "g711_alaw"
• "pcm16"

Default value: "pcm16"

Configuration for input audio transcription, defaults to off and can be set to null to turn off once on. Input audio transcription is not native to the model, since the model consumes audio directly. Transcription runs asynchronously through the /audio/transcriptions endpoint and should be treated as guidance of input audio content rather than precisely what the model heard. The client can optionally set the language and prompt for transcription, these offer additional guidance to the transcription service.

An optional text to guide the model's style or continue a previous audio segment. For whisper-1, the prompt is a list of keywords. For gpt-4o-transcribe models, the prompt is a free text string, for example "expect words related to technology".

The model to use for transcription, current options are gpt-4o-transcribe, gpt-4o-mini-transcribe, and whisper-1.

The language of the input audio. Supplying the input language in ISO-639-1 (e.g. en) format will improve accuracy and latency.

The set of modalities the model can respond with. To disable audio, set this to ["text"].

The default system instructions (i.e. system message) prepended to model calls. This field allows the client to guide the model on desired responses. The model can be instructed on response content and format, (e.g. "be extremely succinct", "act friendly", "here are examples of good responses") and on audio behavior (e.g. "talk quickly", "inject emotion into your voice", "laugh frequently"). The instructions are not guaranteed to be followed by the model, but they provide guidance to the model on the desired behavior.

Note that the server sets default instructions which will be used if this field is not set and are visible in the session.created event at the start of the session.

The format of input audio. Options are pcm16, g711_ulaw, or g711_alaw. For pcm16, input audio must be 16-bit PCM at a 24kHz sample rate, single channel (mono), and little-endian byte order.

Possible values:

• "g711_ulaw"
• "g711_alaw"
• "pcm16"

Default value: "pcm16"