Skip to navigation
Models and Languages

Nova-3 Language Prompting

Use language hints with Nova-3 multilingual to bias detection toward the languages you expect across 55 languages, or let the model auto-detect.

Pre-recorded Streaming:Nova

Nova-3 multilingual (model=nova-3&language=multi) transcribes 55 languages with one model, including speech that switches languages mid-conversation. The optional language_hint parameter biases detection toward the languages you expect. Without hints, the model auto-detects the spoken language.

Nova-3 multilingual uses the same /v1/listen endpoints, model string, and API key you already use for pre-recorded and streaming transcription.

Supported Languages

LanguageCode
Afrikaansaf
Arabicar
Belarusianbe
Bengalibn
Bosnianbs
Bulgarianbg
Catalanca
Chinese (Cantonese, Traditional)zh-HK
Chinese (Mandarin, Simplified)zh, zh-CN, zh-Hans
Croatianhr
Czechcs
Danishda
Dutchnl
Englishen, en-US, en-GB, en-AU, en-NZ, en-IN, en-CA
Estonianet
Finnishfi
Flemishnl-BE
Frenchfr
Galiciangl
Germande
Greekel
Gujaratigu
Hebrewhe
Hindihi
Hungarianhu
Indonesianid
Italianit
Japaneseja
Kannadakn
Koreanko
Latvianlv
Lithuanianlt
Macedonianmk
Malayms
Marathimr
Norwegianno
Persianfa
Polishpl
Portuguesept
Romanianro
Russianru
Serbiansr
Slovaksk
Sloveniansl
Spanishes
Swedishsv
Tagalogtl
Tamilta
Telugute
Thaith
Turkishtr
Ukrainianuk
Vietnamesevi
Xhosaxh
Zuluzu

Locale subtags (for example, en-GB or pt-BR) are accepted as hints, resolve to the base language, and are returned as the base code in languages_hinted.

The language_hint Parameter

language_hint string or array of strings (optional, repeatable)

Pass one or more language_hint values to bias detection toward specific languages. Repeat the parameter to pass several hints, for example language_hint=en&language_hint=es. Hints bias detection; they do not restrict it, so the model still transcribes any supported language it hears.

BehaviorDescription
Single hintBiases toward one language. Use it when you know the language ahead of time.
Multiple hintsBiases toward a set of languages. Use it for bilingual or multilingual audio.
No hintThe model auto-detects the language. Use it when the language is unknown.

language_hint only applies when language=multi; with any other language value it is ignored. Nova-2 does not support language_hint. A hint that is not one of the codes in the Supported Languages table fails the request with HTTP 400 INVALID_QUERY_PARAMETER; for streaming, the WebSocket connection is rejected with HTTP 400.

Set hints per request for pre-recorded audio, or in the connection URL for streaming audio. Hints stay fixed for the life of a streaming connection.

Usage Scenarios

The examples below use the streaming endpoint. The same language_hint values work on pre-recorded requests to https://api.deepgram.com/v1/listen.

1. Known Single Language

When you know the speaker’s language ahead of time (for example, a Spanish-language call center), set a single language_hint.

wss://api.deepgram.com/v1/listen?model=nova-3&language=multi&language_hint=es&encoding=linear16&sample_rate=16000

2. Known Subset of Languages

When speakers may use one of several languages (for example, a bilingual English/Spanish support line), pass multiple hints. The model biases toward the set and still transcribes whichever language is spoken.

wss://api.deepgram.com/v1/listen?model=nova-3&language=multi&language_hint=en&language_hint=es&encoding=linear16&sample_rate=16000

3. Unknown Language

When you don’t know which language the speaker will use, omit language_hint. The model auto-detects the language from the audio.

wss://api.deepgram.com/v1/listen?model=nova-3&language=multi&encoding=linear16&sample_rate=16000

4. Code-Switching

When speakers switch between languages mid-conversation, hint every language you expect, or send no hints. Nova-3 multilingual handles mid-sentence language switches natively.

wss://api.deepgram.com/v1/listen?model=nova-3&language=multi&language_hint=en&language_hint=es&language_hint=fr&encoding=linear16&sample_rate=16000

Language Fields in the Response

With language=multi, each alternative in the response includes these fields:

FieldTypeDescription
languagesstring array (BCP-47)Languages detected in the transcript, sorted by word count (descending). Also returned on each utterance when utterances=true.
languages_hintedstring arrayThe language_hint values for the request or connection, normalized: canonical case (ES → es, zh-hk → zh-HK), and a regional code the model has no tag for becomes its base language (en-GB → en). Returned on every Nova-3 multilingual response; an empty array when no hint was sent.
words[].languagestring (BCP-47)The detected language of each word.

Pre-recorded responses return these fields in results.channels[].alternatives[]. Streaming responses return them in channel.alternatives[] on every interim and final result.

Pre-Recorded Example

cURL
curl \
  --request POST \
  --header 'Authorization: Token YOUR_DEEPGRAM_API_KEY' \
  --header 'Content-Type: audio/wav' \
  --data-binary @youraudio.wav \
  --url 'https://api.deepgram.com/v1/listen?model=nova-3&language=multi&language_hint=en&language_hint=es'

The response, trimmed to the language fields:

{
  "results": {
    "channels": [
      {
        "alternatives": [
          {
            "transcript": "después de revisar su póliza de automóvil veo que actualmente cuenta con cobertura integral con un deducible de quinientos dólares",
            "confidence": 1.0,
            "languages": ["es"],
            "languages_hinted": ["en", "es"],
            "words": [
              { "word": "después", "start": 0.32, "end": 0.79999995, "confidence": 1.0, "language": "es" },
              { "word": "de", "start": 0.79999995, "end": 0.88, "confidence": 0.99902344, "language": "es" },
              { "word": "revisar", "start": 0.88, "end": 1.36, "confidence": 1.0, "language": "es" },
              { "word": "su", "start": 1.36, "end": 1.52, "confidence": 1.0, "language": "es" },
              { "word": "póliza", "start": 1.52, "end": 2.0, "confidence": 1.0, "language": "es" },
              { "word": "de", "start": 2.0, "end": 2.1599998, "confidence": 0.99902344, "language": "es" },
              { "word": "automóvil", "start": 2.1599998, "end": 2.96, "confidence": 0.9958008, "language": "es" },
              { "word": "veo", "start": 2.96, "end": 3.28, "confidence": 0.9995117, "language": "es" },
              { "word": "que", "start": 3.28, "end": 3.4399998, "confidence": 1.0, "language": "es" },
              { "word": "actualmente", "start": 3.4399998, "end": 4.0, "confidence": 1.0, "language": "es" },
              { "word": "cuenta", "start": 4.0, "end": 4.3199997, "confidence": 1.0, "language": "es" },
              { "word": "con", "start": 4.3199997, "end": 4.56, "confidence": 1.0, "language": "es" },
              { "word": "cobertura", "start": 4.56, "end": 5.12, "confidence": 1.0, "language": "es" },
              { "word": "integral", "start": 5.12, "end": 5.7599998, "confidence": 1.0, "language": "es" },
              { "word": "con", "start": 5.7599998, "end": 6.0, "confidence": 0.9550781, "language": "es" },
              { "word": "un", "start": 6.0, "end": 6.16, "confidence": 0.99902344, "language": "es" },
              { "word": "deducible", "start": 6.16, "end": 6.72, "confidence": 0.9482422, "language": "es" },
              { "word": "de", "start": 6.72, "end": 6.8799996, "confidence": 1.0, "language": "es" },
              { "word": "quinientos", "start": 6.8799996, "end": 7.3599997, "confidence": 0.9946289, "language": "es" },
              { "word": "dólares", "start": 7.3599997, "end": 9.36, "confidence": 0.9970703, "language": "es" }
            ]
          }
        ]
      }
    ]
  }
}

Streaming Example

Open a WebSocket connection with hints in the URL:

wss://api.deepgram.com/v1/listen?model=nova-3&language=multi&language_hint=en&language_hint=es&encoding=linear16&sample_rate=16000

Each Results message, interim and final, includes the language fields:

{
  "type": "Results",
  "channel_index": [0, 1],
  "duration": 2.91,
  "start": 2.84,
  "is_final": true,
  "speech_final": true,
  "channel": {
    "alternatives": [
      {
        "transcript": "veo que actualmente cuenta con cobertura integral",
        "confidence": 1.0,
        "languages": ["es"],
        "languages_hinted": ["en", "es"],
        "words": [
          { "word": "veo", "start": 2.9199998, "end": 3.32, "confidence": 0.97314453, "language": "es" },
          { "word": "que", "start": 3.32, "end": 3.3999999, "confidence": 1.0, "language": "es" },
          { "word": "actualmente", "start": 3.3999999, "end": 4.04, "confidence": 1.0, "language": "es" },
          { "word": "cuenta", "start": 4.04, "end": 4.3599997, "confidence": 1.0, "language": "es" },
          { "word": "con", "start": 4.3599997, "end": 4.52, "confidence": 1.0, "language": "es" },
          { "word": "cobertura", "start": 4.52, "end": 5.08, "confidence": 1.0, "language": "es" },
          { "word": "integral", "start": 5.08, "end": 5.4799995, "confidence": 1.0, "language": "es" }
        ]
      }
    ]
  }
}

Use the languages field to route downstream processing, for example to pick the TTS voice or LLM prompt language that matches what the speaker said.