Aller au contenu principal

Speech to Text

Category: AI & ML
Node id: speech-to-text

Standards: HIPAA

Transcribes audio files to text using ElevenLabs Scribe v1 (speaker diarization, 99 languages, timestamped segments) or Gemini on Vertex AI (provider: gemini)

What it does​

  • Audio transcription via ElevenLabs Scribe v1 (state-of-the-art accuracy)
  • 99 language support with automatic language detection
  • Speaker diarization with configurable speaker count
  • Timestamped segments at word or sentence level
  • Supports audio URLs or inline audio data
  • Handles files up to 3GB / 10 hours duration

Use cases​

  • Transcribe physician-patient encounter recordings for SOAP note generation
  • Convert telemedicine audio to searchable clinical documentation
  • Process Docline consultation recordings for clinical workflow pipelines
  • Multi-speaker clinical meeting transcription with speaker identification

Runtime​

  • Performance tier: slow

Input ports​

NameTypeRequiredDescription
audio_datatextNoBase64-encoded or raw audio data
audio_urltextNoURL to fetch audio file from

Output ports​

NameTypeRequiredDescription
transcripttextNoFull transcription text
segmentsjsonNoTimestamped segments with speaker labels
speakersjsonNoSpeaker information from diarization
languagetextNoDetected or specified language code
durationnumberNoAudio duration in seconds
successbooleanNoWhether transcription succeeded

Settings​

Basic​

SettingNameTypeRequiredDefaultAllowed valuesDescription
ProviderproviderenumNoelevenlabselevenlabs, geminiTranscription provider: ElevenLabs Scribe, or Gemini on Vertex AI (Application Default Credentials, no API key)
Gemini ModelmodelstringNoGemini model for provider gemini (falls back to GEMINI_MODEL)
Modelmodel_idenumNoscribe_v1scribe_v1, scribe_v1_experimentalTranscription model
LanguagelanguagestringNoLanguage code (auto-detect if empty)
Detect Speakersdetect_speakersbooleanNotrueEnable speaker diarization
Number of Speakersnum_speakersnumberNo0min 0; max 20Expected speaker count (0 = auto)

Auth​

SettingNameTypeRequiredDefaultAllowed valuesDescription
ElevenLabs API Keyapi_keysecretNoAPI key for ElevenLabs (provider elevenlabs only)
GCP Projectgcp_projectstringNoVertex AI project for provider gemini (falls back to GEMINI_PROJECT / GOOGLE_CLOUD_PROJECT)
GCP Locationgcp_locationstringNoVertex AI location where the recording is processed, provider gemini (falls back to GEMINI_LOCATION; no implicit global default)

Advanced​

SettingNameTypeRequiredDefaultAllowed valuesDescription
Audio MIME Typemime_typestringNoAudio type sent to Gemini (e.g. audio/wav, audio/mpeg); detected from the input when empty
Vertex Base URLbase_urlstringNoVertex AI endpoint override for provider gemini (private endpoints, testing)
API URLapi_urlstringNohttps://api.elevenlabs.io/v1/speech-to-textElevenLabs STT endpoint
Timestamp Granularitytimestamp_granularityenumNosegmentsegment, wordTimestamp level
Timeout (seconds)timeoutnumberNo300min 10; max 600Request timeout for large files

Errors​

CodeDescription
API_KEY_REQUIREDElevenLabs API key not configured (provider elevenlabs)
VERTEX_NOT_CONFIGUREDNo Vertex AI project, location or model for provider gemini (gcp_* / model config or GEMINI_* env)
NO_AUDIO_INPUTNeither audio_data nor audio_url was provided
AUDIO_FETCH_FAILEDFailed to download audio from the provided URL
STT_REQUEST_FAILEDElevenLabs API call failed
EMPTY_TRANSCRIPTTranscription succeeded but returned no text

Runtime notes​

  • Provide either audio_data (base64/raw bytes) or audio_url (fetched automatically) — not both
  • transcript is the concatenated full text; segments contains per-utterance timestamps and speaker IDs
  • Speaker diarization is enabled by default; set detect_speakers=false to skip
  • Timeout defaults to 300s (5 min) to accommodate large audio files up to 3GB / 10 hours
  • success=false with an empty transcript when the audio contains no recognizable speech

See Build any workflow for how to put nodes together and Ports, types and data flow for how port types connect.

Quelque chose n'est pas clair ? Envoyez vos commentaires sur la documentation par courriel.