Transcribe

The Transcribe node turns speech into text. Wire in audio or a video, and the node returns the transcript twice: as plain text, and as a structured transcript with the timing of every word. You can choose between four engines, and they differ in one thing that matters for captions: whether they return word timings. Nemotron 3.5 Transcribe can also translate the speech instead of transcribing it.

When to use it

• Transcribe an interview or a podcast for written content.
• Make captions and subtitles from the sound of a video, with Add Captions.
• Turn voice memos and meeting recordings into text.
• Make an audio library searchable as text.
• Feed a transcript into a text node to summarize or analyze it.
• Give Edit Plan the timed transcript it needs to plan an edit.

Quick start

Input and outputs

To transcribe a video from a link to where it is, connect a Video URL node. The Video URL node asks for the content rights declaration and fetches the sound, and Transcribe transcribes that track. A link to a video's page given to Transcribe in another way stops a run in the editor, and the message says so.

In a prompt, a reference to the node's label, such as {Transcribe}, gives the plain text.

Settings

Speaker Diarization and Tag Audio Events are independent: turn on one, both or neither.

Models

Word timings: which engine returns them

If ElevenLabs STT, Incredibly Fast Whisper or Nemotron 3.5 Transcribe returns no word timings for audio that clearly has speech, the run fails and the credits are refunded. Audio with no speech at all still succeeds, with no words, for example a clip that transcribes as [music].

Translate the speech

With Translate to set, Nemotron 3.5 Transcribe translates the speech instead of transcribing it. Translation goes to or from English:

• Into English. Speech in another language becomes English text.
• From English. English speech becomes Chinese, Japanese, Korean, French, German, Spanish, Portuguese, Italian, Russian, Dutch, Polish, Czech, Slovak, Hindi, Bulgarian, Croatian, Danish, Estonian, Finnish, Hungarian, Latvian, Lithuanian, Romanian, Slovenian, Swedish, Thai, Ukrainian, Arabic, Greek, Hebrew or Maltese text.

A translation between two other languages may work, but it is not guaranteed. For the best result, translate into English.

• The Text output carries the translation, and the transcript's language is the target language.
• A translation has no word timings, so the words list of the Transcript output is empty. A translating node that feeds Add Captions is refused before the run.
• Only Nemotron 3.5 Transcribe translates. A translation costs the same as a transcript.

The file Nemotron 3.5 Transcribe takes

Nemotron 3.5 Transcribe takes a file of up to 60 minutes and 200 MB. The file is MP3, WAV, M4A or AAC, FLAC, OGG or Opus, or WEBM audio, or an MP4 or MOV video with a sound track. It must be a direct HTTPS link to a file on a public host.

• Measured first. The node measures the file before it holds any credits, because the price depends on the length.
• Refused with nothing charged: a link that is not HTTPS to a public host, a file over 60 minutes or 200 MB, and a link to a video's page on a social platform. For such a video, bring its sound in through a Video URL node first, or choose another engine.
• A length that cannot be read up front is not refused. A browser recording saved as a stream or a fragmented MP4 holds the 60-minute price, and the run is charged for the length the model decoded.
• Speakers. With word timings, voices that sound very different separate well. Similar voices may be merged into one speaker.
• Repeated text. When the answer reaches the model's length limit for the audio, the job's output carries truncated: true. This usually means that the model repeats itself, so review or discard the result.

The structured transcript

The Transcript output looks like this:

{
  "version": 1,
  "language": "en",
  "words": [
    { "text": "Welcome", "startMs": 0, "endMs": 420 },
    { "text": "back", "startMs": 460, "endMs": 700 }
  ],
  "segments": [
    { "startMs": 0, "endMs": 700, "text": "Welcome back" }
  ]
}

• All times are whole milliseconds.
• words holds every word with its start and end. It is filled whenever the engine returns word timings.
• segments holds the coarser sentences or phrases, when the engine provides them.
• On a run with Speaker Diarization, each word and segment also has a speaker. A word can also carry a confidence.

On the Text output, Speaker Diarization adds labels such as Speaker 1: before each segment, and Tag Audio Events adds tags such as [music] and [laughter] in place.

Make captions from a transcript

Wire the Transcript output into the transcript input of Add Captions. Add Captions then draws captions timed to each word. Use ElevenLabs STT, Incredibly Fast Whisper or Nemotron 3.5 Transcribe for this.

A Whisper transcript cannot drive captions. When a Transcribe node on Whisper feeds Add Captions, the run is refused before anything runs or is charged, with this message:

Captions need word timings, but the "whisper" engine does not return word timings — pick incredibly-fast-whisper or elevenlabs-stt.

• A Nemotron 3.5 Transcribe node set to translate is refused the same way, because a translation has no word timings. Set Translate to to Don't translate to run it.
• The check also covers a chain through Apply EDL, and chains inside a sub-workflow at any depth.
• The settings panel shows the same warning next to Provider as soon as the connection exists, so you can fix it before you run.
• Running the Transcribe node on its own is never blocked. A Whisper transcript that goes anywhere else is fine.
• One case cannot be checked in advance: a chain where Transcribe and Add Captions sit on different sides of a sub-workflow boundary. That run fails at Add Captions, after the transcription has run and been charged.

Credits

Three engines cost a flat price per run, whatever the length of the audio:

Nemotron 3.5 Transcribe is priced by the length of the file. It charges per started minute of the audio, up to 60 minutes, and the price is the same with word timings or a translation. The minutes are added up, and the run is rounded up to a whole credit once.

• Before a run, the Run button shows the price of the length the workflow knows, such as the length of an upload. When no length is known yet, it shows the 60-minute price.
• The hold. The file is measured before the run, and the price of that length is held. A file whose length cannot be read holds the 60-minute price.
• The charge. After the run, the run is charged for the length the model actually decoded. This can differ from the length in the file's header, because some MP3 files state a wrong length. A shorter length releases the rest of the hold, and a longer one takes the difference from your balance.

A request that is refused before it starts, such as a Whisper transcript wired into Add Captions, is not charged.

On a self-hosted install

Each engine runs on your own provider key. When the chosen engine has no key and the install has a Nodaro Cloud connection, the transcription runs through the connection. With neither, the node fails with a message that names the key to add. See Provider keys.

Tips

• Leave the language on Auto Detect. Choose a language only when you know it, for example when two languages sound similar.
• Label speakers for conversations. Turn on Speaker Diarization for interviews, meetings and podcasts.
• Tag events when they matter. Turn on Tag Audio Events when background sounds help to understand the recording.
• Clean noisy audio first. Wire the recording through Voice Extractor before Transcribe.
• Split very long audio. Shorter segments transcribe more reliably. Split into Chunks cuts a long file into parts between a minimum and a maximum length, and Transcribe runs once for every part.
• Align your own script instead. When you already have the exact script, Forced Alignment times each of its words against the audio.

From the API

POST /v1/transcribe takes audioUrl and the optional provider (elevenlabs-stt, incredibly-fast-whisper, whisper or nemotron-3-5-transcribe), language, diarize, tagAudioEvents, wordTimestamps and targetLanguage.

• A link to a video's page. An audioUrl that links to a video's page on a social platform is an import of that link: its sound is fetched first. Send the confirmationId from POST /v1/content-rights/confirm for that link. Without one, a deployment that requires a confirmation refuses the request with 403 before the job is created. A key or an agent whose account holds a standing authorization needs none, and a direct file link needs none. See Content rights.

• Name the engine. A request without provider runs Whisper, so existing callers keep the engine they had. Set provider whenever you need word timings, diarize or tagAudioEvents.

• Word timings. incredibly-fast-whisper and nemotron-3-5-transcribe return words only with wordTimestamps: true. wordTimestamps: true on whisper is refused with a 400 validation_error before the job is created, so nothing is charged. The engine is never swapped silently.

• Translation. targetLanguage works only on nemotron-3-5-transcribe. On another engine, in a language it does not translate into, or together with wordTimestamps: true, it is refused with a 400 validation_error on targetLanguage, and nothing is charged.

• Captions. The finished job's output_data.words already has the caption shape (text, startMs, endMs). Pass it to POST /v1/add-captions as captions, and Add Captions does not transcribe again.

• Price. GET /v1/credits/model-cost?model=<engine> returns the exact price for your account. For Nemotron 3.5 Transcribe, ask for nemotron-3-5-transcribe:<minutes>m.

The SDK method is client.audio.transcribe, and the CLI command is nodaro audio transcribe --provider <engine>, with --target-language for a translation. The MCP tool transcribe takes the engine in model and a translation in target_language. Without model it runs ElevenLabs STT, so its result has word timings. See Voice and media.