# Speech understanding Speech understanding enables you to configure recognition, choose a model, and check language support. Put recognition settings in the `config` object for Fast and Batch requests or in `config` within a Live `session.update` event. ## Recognition configuration | Field | Type | Default | Fast | Batch | Live | Description | | -------------------- | ------- | --------- | ---- | ----- | --------------- | ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | | `language` | string | `en-US` | Yes | Yes | Yes | Recognition language. With `language_detection`, the preferred initial language; when `language_detection` is set and `language` is omitted, there is no preferred language. | | `model` | string | Automatic | Yes | Yes | Yes | Recognition model. Omit when `language_detection.mode` is `once`; Scribe chooses the model from the detected language. | | `language_detection` | object | — | Yes | Yes | Continuous only | Detects a language once or continuously. Live supports only `continuous`. | | `interim_results` | boolean | — | No | No | Yes | Enables Captions: partial transcription results before a transcript item is completed. | | `segmentation_mode` | string | `auto` | Yes | Yes | No | Use `auto` to split audio into segments at pauses in speech (voice activity detection), or `none` to disable segmentation. | | `word_time_offsets` | boolean | `false` | Yes | Yes | No | Includes word-level start and end offsets. | | `channel_separation` | boolean | `false` | Yes | Yes | No | Processes stereo channels independently. | | `diarization` | boolean | `false` | Yes | Yes | Yes | Identifies speakers. When `false`, `speaker` is an empty string. | | `profanity_filter` | boolean | `false` | Yes | Yes | Yes | Filters profanity when the selected model supports it. | | `output_format` | string | `json` | Yes | Yes | No | Output format. Defaults to `json`. | | `vocabulary` | object | — | Yes | Yes | Yes | Inline vocabulary for the current request, job, or session. | | `vocabulary_id` | string | — | Yes | Yes | Yes | Persistent Vocabulary Asset to apply. | Unsupported combinations of language, model, execution mode, and settings return a validation error. ## Select a model Set `config.model` to explicitly select a recognition model. When it is omitted, Scribe selects a compatible standard model from the language, detection mode, and execution mode. When `config.language_detection.mode` is `once`, omit `config.model`. Combining an explicit model with `once` returns HTTP `400` because Scribe must select the model after detecting the language. | Model | Coverage | Language detection | | -------------------------- | -------------------------------------------------------------------------------- | ------------------ | | `zoom-scribe-en` | English | No | | `zoom-scribe-en-pro` | English | No | | `zoom-scribe-multilingual` | Ten languages; see [Multilingual model languages](#multilingual-model-languages) | Continuous | | `zoom-scribe-` | A supported language-specific model | No | | `whisper` | 36 languages, including 26 locales not covered by `zoom-scribe-multilingual` | Continuous | The multilingual model IDs are the same in Fast and Live. Set `config.model` to the identifier below. You do not need a Live-specific prefix. | Model | Fast `config.model` | Live `config.model` | | ----------------- | -------------------------- | -------------------------- | | Zoom multilingual | `zoom-scribe-multilingual` | `zoom-scribe-multilingual` | | Whisper | `whisper` | `whisper` | Models ending in `-pro` require premium-model access. Automatic selection does not choose a premium model unless you enable premium usage for the account. `zoom-scribe-en-pro` is the Pro option for English, **Scribe API 3B Pro High-Accuracy**, for applications that need higher accuracy. Use the model ID, not the product name, in `config.model`. The response's top-level `model` field identifies the model that processed the input. ```json { "file": "https://example.com/audio.wav", "config": { "language": "en-US", "model": "zoom-scribe-en", "word_time_offsets": true, "diarization": true } } ``` ## Detect languages Use `config.language_detection` to detect the input language. | Mode | Behavior | | ------------ | ------------------------------------------------------------------------------------------------ | | `once` | Detects one language, then recognizes the rest of the input with that language. | | `continuous` | Detects language changes as the input is processed. Each segment contains one detected language. | Live supports only `continuous`. For `once`, omit `model` so Scribe can select it based on the detected language: ```json { "config": { "language_detection": { "mode": "once", "candidate_languages": ["en-US", "fr-FR"] } } } ``` Use `candidate_languages` to restrict detection. When both `language` and `candidate_languages` are present, the preferred `language` must be included in the candidate list. For `continuous`, you can explicitly select a model that supports continuous detection: ```json { "config": { "language": "en-US", "model": "zoom-scribe-multilingual", "language_detection": { "mode": "continuous", "candidate_languages": ["en-US", "es-ES", "fr-FR"] } } } ``` For Fast and Batch continuous detection, `result.language` and `result.segments[].language` are omitted because the input can contain multiple languages. Do not expect per-segment language labels. With `once`, `result.language` may also be absent; clients should handle its omission. Live continuous detection returns the detected language in `transcription.completed.language`, such as `fr` for French. ## Multilingual model languages The `zoom-scribe-multilingual` model supports these languages and locales: | Language | Locale codes | | -------------------- | ---------------- | | English | `en-US` | | Arabic | `ar-SA`, `ar-AE` | | Chinese (Simplified) | `zh-CN` | | French | `fr-FR` | | German | `de-DE` | | Italian | `it-IT` | | Japanese | `ja-JP` | | Korean | `ko-KR` | | Portuguese | `pt-BR`, `pt-PT` | | Spanish | `es-ES` | ## Broader language coverage The `whisper` model recognizes 36 languages, with integrated language detection. These include the ten languages of the multilingual model and the following 26 locales: | Language | Locale code | Language | Locale code | | ---------- | ----------- | ------------------ | ----------- | | Bengali | `bn-IN` | Malay | `ms-MY` | | Cantonese | `zh-HK` | Persian (Farsi) | `fa-IR` | | Catalan | `ca-ES` | Polish | `pl-PL` | | Czech | `cs-CZ` | Romanian | `ro-RO` | | Danish | `da-DK` | Russian | `ru-RU` | | Dutch | `nl-NL` | Swedish | `sv-SE` | | Estonian | `et-EE` | Tagalog (Filipino) | `tl-PH` | | Finnish | `fi-FI` | Tamil | `ta-IN` | | Greek | `el-GR` | Telugu | `te-IN` | | Hebrew | `he-IL` | Thai | `th-TH` | | Hindi | `hi-IN` | Turkish | `tr-TR` | | Hungarian | `hu-HU` | Ukrainian | `uk-UA` | | Indonesian | `id-ID` | Vietnamese | `vi-VN` | ## Transcript response Fast returns the result inline. Batch writes the same result schema to the configured storage destination. ```json { "request_id": "3f2b8c1e-7a4d-4e9b-b6f0-1c5d8a2e9b47", "duration_sec": 10.99, "result": { "text_display": "Welcome everyone. Let's get started.", "segments": [ { "id": "utt_0000", "start": 0.0, "end": 2.4, "channel": 0, "speaker": "Speaker 1", "text": "Welcome everyone.", "words": [{ "word": "Welcome", "start": 0.0, "end": 0.4 }] } ] }, "model": "zoom-scribe-en", "usage": { "input_units": 10.99, "output_units": 0, "unit_type": "seconds" } } ``` ## Process multiple channels Fast and Batch accept mono and stereo media. Set `channel_separation` to `true` to recognize each stereo channel independently. ```json { "config": { "language": "en-US", "channel_separation": true, "diarization": false } } ``` Every segment includes `channel`. Channel `0` is the left channel and channel `1` is the right channel. When separation is disabled or the input is mono, the value is `0`.