Translation Settings Breakdown

A complete reference for all configuration options of the set_task message.

What are translation settings?

Translation settings control how your audio is processed, transcribed, translated, and synthesized. They are included in the data field of a set_task message sent via the WebRTC data channel or the WebSocket connection.

You will use these settings when:

Starting a new translation — send your initial set_task message to begin real-time translation.
Updating an active translation — modify settings during an ongoing translation without interrupting the stream.
Fine-tuning performance — adjust parameters to optimize quality, latency, or specific use cases.

For ready-to-use values, see Recommended settings.

Message envelope

{
  "message_type": "set_task",
  "data": {
    "input_stream":  { /* ... */ },
    "output_stream": { /* ... */ },
    "pipeline":      { /* ... */ }
  }
}

`message_type`

Type: string
Description: Identifies the type of command being sent:
- "set_task" — create a new task or update an existing one.
- "get_task" — return the current task.
- "pause_task" — stop processing new audio but keep the task alive. Use set_task to resume.
- "interrupt_task" — instantly stop the speech of the current (ongoing) phrase.
- "end_task" — end the task (you will be disconnected automatically).
- "tts_task" — generate text-to-speech from text (temporarily disabled; use the Realtime TTS API).
- "input_audio_data" — WebSocket transport only; contains an audio data chunk.

`data`

Type: object
Description: The main task configuration. Contains three sections:

The translation_queue_configs and allowed_message_types fields belong inside pipeline.

`input_stream`

Configures the input stream.

Field	Type	Description
`content_type`	string	Content type of the input stream. Currently only `audio` is supported.
`source`	object	How and from where the input audio is sourced. See below.

`source`

Field	Type	Description
`type`	string	Input audio transport: `webrtc` or `ws`. Must match the `target` transport in `output_stream`.
`format`	string	WebSocket transport only. Input audio format: `opus`, `pcm_s16le`, or `wav`.
`sample_rate`	integer	WebSocket transport only. Input sample rate. Allowed range: 16000–48000 Hz.
`channels`	integer	WebSocket transport only. Number of input channels: 1 or 2.

`output_stream`

Configures the output stream.

Field	Type	Description
`content_type`	string	Content type of the output. Currently only `audio` is supported.
`target`	object	Destination of the output. See below.

`target`

Field	Type	Description
`type`	string	Output audio transport: `webrtc` or `ws`. Must match the `source` transport in `input_stream`.
`format`	string	WebSocket transport only. Output audio format: `pcm_s16le` or `zlib_pcm_s16le` (zlib-compressed PCM).
`sample_rate`	integer	WebSocket transport only. Output sample rate. Fixed at 24000 Hz.
`channels`	integer	WebSocket transport only. Output channels. Fixed at 1 (mono).

`pipeline`

Holds the configuration for all processing steps: preprocessing, transcription, and translation.

`transcription`

Settings for Automatic Speech Recognition (ASR).

`source_language`

Type: string
Description: Language code of the input audio (e.g., "en", "es", "fr"). Set to "auto" to enable automatic language detection (optionally restricted by detectable_languages). See Supported languages.

`detectable_languages`

Type: array of strings
Description: When source_language is "auto", only languages from this list will be detected. Leave empty to allow any supported language.

`segment_confirmation_silence_threshold`

Type: float
Description: Seconds of silence needed to confirm the end of a segment (0.3–2.0, default 0.7). Recommended range: 0.5–0.9 s, depending on the speaker's tempo and pauses. Increase it if the speaker frequently pauses between words; setting it too low can cause unwanted sentence splitting.

`speakers_total`

Type: integer or null
Description: Expected number of speakers (1–1000). Helps speaker handling when known in advance.

`only_confirm_by_silence`

Type: bool
Description: When true, segments are confirmed only by silence detection.

`sentence_splitter`

Type: object
Description: Controls how longer sentences are split into smaller parts (sometimes with slight rephrasing, without losing the meaning) to speed up processing.

Field	Type	Description
`enabled`	bool	Whether to enable automatic sentence splitting.

`allow_hotwords_glossaries`

Type: bool
Default: true
Description: Applies your Recognition (asr_hot) glossaries for the source language: terms the speech recognition should get right. See Glossary settings.

`allowed_hotwords_glossary_ids`

Type: array of strings or null
Default: null
Description: IDs of the Recognition glossaries to apply (1–15). If not set, all your enabled Recognition glossaries for the source language apply. See Glossary settings.

`verification`

Type: object
Description: Transcription verification settings.

Field	Type	Description
`allow_verification_glossaries`	bool	Applies your Validation (`asr`) glossaries for the source language: recognized text is replaced with your terms. Default `true`. See Glossary settings.
`allowed_verification_glossary_ids`	array of strings or null	IDs of the Validation glossaries to apply (1–15). If not set, all your enabled Validation glossaries for the source language apply.
`auto_transcription_correction`	bool	WIP. Enables automatic transcription verification using an LLM.
`transcription_correction_style`	string or null	Style of the LLM correction.

`translations`

Type: array of objects
Description: An array of translation targets. Each object defines translation settings for one target language; add one object per target language.

`target_language`

Type: string
Description: Language to translate into (e.g., "en-us", "es", "fr"). See Supported languages.

`allowed_source_languages`

Type: array of strings
Description: Restricts this translation target to specific source languages. Used for conditional multi-language translation together with "source_language": "auto".

`translate_partial_transcriptions`

Type: bool
Description: Enables translation of partial (unconfirmed) transcriptions.

`allow_translation_glossaries`

Type: bool
Default: true
Description: Applies your Translation (translation) glossaries for this source → target language pair. See Glossary settings.

`allowed_translation_glossary_ids`

Type: array of strings or null
Default: null
Description: IDs of the Translation glossaries to apply to this target (1–15). If not set, all your enabled Translation glossaries for the language pair apply. See Glossary settings.

`speech_generation`

Type: object
Description: Text-to-speech (TTS) settings for this target language.

Field	Type	Description
`voice_cloning`	bool	Experimental. Mimics the original speaker's voice. It usually takes 10–20 seconds of speech before the voice changes are applied.
`voice_id`	string or null	A specific voice ID (voice cloning must be disabled). `"default_low"` or `"default_high"` automatically picks the best default voice for the language. Manage voices in the Palabra web portal.
`voice_timbre_detection`	object	Automatically detects voice timbre and assigns voice IDs accordingly. See below.

`voice_timbre_detection`

Field	Type	Description
`enabled`	bool	Enables voice timbre detection (voice cloning must be disabled).
`high_timbre_voices`	array of strings	Voice ID to use for high-timbre voices. Currently only one ID is supported; `"default_high"` can be used.
`low_timbre_voices`	array of strings	Voice ID to use for low-timbre voices. Currently only one ID is supported; `"default_low"` can be used.

Glossary settings

You create and enable glossaries with the Glossaries API or in the web portal. The fields below choose which of them a task applies. Each glossary type has its own pair of fields:

Glossary type	`glossary_type`	Fields	Matched by
Recognition	`asr_hot`	`pipeline.transcription.allow_hotwords_glossaries`, `pipeline.transcription.allowed_hotwords_glossary_ids`	source language
Validation	`asr`	`pipeline.transcription.verification.allow_verification_glossaries`, `pipeline.transcription.verification.allowed_verification_glossary_ids`	source language
Translation	`translation`	`pipeline.translations[N].allow_translation_glossaries`, `pipeline.translations[N].allowed_translation_glossary_ids`	source → target language pair of that target

The rules are the same for every type:

Nothing set (default): all your enabled glossaries of the type that match the language apply.
allow_*_glossaries: false: no glossaries of the type apply.
allowed_*_glossary_ids: only the listed glossaries apply. The list takes 1–15 unique IDs. An empty list makes the whole set_task fail with VALIDATION_ERROR (a running task keeps its previous settings). To apply no glossaries, use allow_*_glossaries: false.
Only enabled glossaries are applied. A listed ID that is disabled, deleted, unknown or for another language is skipped without an error.
Glossaries are loaded when the task starts. A later set_task can change which of the loaded glossaries apply. Changes to the glossaries themselves (new, enabled, disabled or deleted ones) take effect in the next session.
Every set_task replaces the whole task, so a field left out of an update goes back to its default.
get_task returns these fields as you set them. It does not list the glossaries that are actually applied.

For example, this task applies one Recognition glossary, no Validation glossaries, one Translation glossary for Spanish and none for French:

"pipeline": {
  "transcription": {
    "source_language": "en",
    "allowed_hotwords_glossary_ids": ["<recognition_glossary_id>"],
    "verification": {
      "allow_verification_glossaries": false
    }
  },
  "translations": [
    {
      "target_language": "es",
      "allowed_translation_glossary_ids": ["<en_es_glossary_id>"]
    },
    {
      "target_language": "fr",
      "allow_translation_glossaries": false
    }
  ]
}

Keeping customers apart

If one account serves several end customers, set the glossary fields for every type and every target in every set_task: list the customer's glossary IDs, or set allow_*_glossaries: false. A field you leave out falls back to all enabled glossaries of the account.

`translation_queue_configs`

Configures the behavior of unspoken TTS buffers.

The global key holds the default settings. You can add language-specific overrides by using the language code as a key (for example, "es" for Spanish).

Field	Type	Description
`desired_queue_level_ms`	integer	Desired average TTS buffer size in milliseconds (2000–163840). With `auto_tempo` enabled, the system tries to keep the buffer at this level. Recommended: 5000–10000 ms.
`max_queue_level_ms`	integer	Maximum TTS queue size in milliseconds (3000–163840). If the queue grows beyond this limit, it is reduced to `desired_queue_level_ms` by dropping older queued audio. Must be greater than `desired_queue_level_ms`; should be at least 2–3× larger.
`auto_tempo`	bool	Automatically corrects speech tempo based on the queue state. Recommended to keep on.
`auto_tempo_max_delay_ms`	integer	Maximum buffer delay in milliseconds for auto tempo (60–10000, default 250).
`min_tempo`	float	Minimum allowed speech speed (1.0–2.0, default 1.0).
`max_tempo`	float	Maximum allowed speech speed (1.0–2.0, default 1.35). Must be ≥ `min_tempo`.

If you don't provide translation_queue_configs, the server applies a default global config: desired_queue_level_ms: 5000, max_queue_level_ms: 20000, min_tempo: 1.15, max_tempo: 1.45, auto_tempo: true.

`allowed_message_types`

This list selects the optional event messages to send. It does not filter audio output, command responses (task_status, current_task, error), or end_of_stream. task_status is not a valid value for this setting.

Type: array of strings
Default: ["translated_transcription", "partial_transcription", "validated_transcription", "turn_changed"]
Description: Selects optional events delivered over the WebSocket or WebRTC data channel.
- "partial_transcription" — emitted for partial transcription segments as they are recognized.
- "partial_translated_transcription" — emitted for partial translated transcriptions if translate_partial_transcriptions is enabled.
- "validated_transcription" — emitted when a transcription segment is fully confirmed.
- "translated_transcription" — emitted when a transcription segment has been translated.
- "turn_changed" — emitted when enabled turn detection identifies a new speaker turn.

What are translation settings?​

Message envelope​

message_type​

data​

input_stream​

source​

output_stream​

target​

pipeline​

transcription​

source_language​

detectable_languages​

segment_confirmation_silence_threshold​

speakers_total​

only_confirm_by_silence​

sentence_splitter​

allow_hotwords_glossaries​

allowed_hotwords_glossary_ids​

verification​

translations​

target_language​

allowed_source_languages​

translate_partial_transcriptions​

allow_translation_glossaries​

allowed_translation_glossary_ids​

speech_generation​

voice_timbre_detection​

Glossary settings​

translation_queue_configs​

allowed_message_types​