Deepgram: Nova-3

Nova-3 is Deepgram’s general-purpose speech-recognition model for converting audio into text. The producer recommends it for meetings, event captions, multi-speaker recordings and challenging conditions such as noise or distant microphones, using either pre-recorded or streaming transcription. Its useful result is recognized speech with timing and confidence information that remains connected to the source audio.

Choose the language workflow

The documented default language is English. A monolingual request selects the spoken language or supported locale. The current general-model table includes English regional variants, Spanish, French, German, Russian, Portuguese, Japanese, Italian, Dutch, Arabic, Chinese scripts and many other languages, including Kazakh. Select from the producer’s supported codes rather than assuming that any locale name is accepted.

The multilingual setting is a separate code-switching workflow. The current Nova-3 table lists ten languages for it: English, Spanish, French, German, Hindi, Russian, Portuguese, Japanese, Italian and Dutch. This set is not identical to the larger monolingual list. For a recording that alternates between supported languages, prepare the multilingual mode instead of forcing the whole conversation into an unrelated single-language setting. Preserve what was spoken before preparing any translated or summarized version.

Recorded audio and transcript preparation

Deepgram’s file workflow accepts uploaded audio bytes or a reference to an audio file. Save the returned transcript when the request completes; the producer does not store transcripts for later retrieval. Common documented input formats and encodings include MP3, MP4, MP2, AAC, WAV, FLAC, PCM, M4A, Ogg, Opus and WebM. Preserve the actual format and channel information with the recording. Give the recording a stable identifier and keep the result associated with the correct source and its channel information. For a meeting, review participant names and technical terms while listening to the original sound. For a lecture or event, retain the sequence of the audio so a later caption or search result can locate the relevant passage.

The producer’s pre-recorded guide lists a maximum file size of 2 GB and recommends extracting the audio stream from a large video file. It also documents a ten-minute processing timeout for Nova requests. That timeout concerns how long the request processes, not a ten-minute maximum recording duration. A useful result includes the ending of the intended recording; a partially retained transcript should not be treated as the complete text.

Read the words, times and confidence

The documented JSON response includes transcript alternatives and recognized words, with each word’s start and end time in seconds from the beginning of the audio and a confidence value. For a multi-speaker transcript, the producer’s optional diarization assigns a speaker label to each word. Pre-recorded diarization also returns speaker confidence; streaming diarization returns the speaker label. Review those labels against the voices in the recording. Overall transcript confidence is also reported on a 0–1 scale. When smart formatting is requested, the result can include punctuation and capitalization as punctuated words, along with readable paragraph information.

Use those fields according to their meaning. A word time locates recognized speech, while confidence is the model’s indication of reliability, not proof that a difficult name or quantity was heard correctly. For caption preparation, retain the timing alongside reviewed wording and inspect the boundaries against the sound. For a readable editorial version, keep changes distinguishable from the recognized transcript rather than silently replacing uncertain words with a plausible story.

Streaming recognition and final segments

Nova’s streaming workflow returns text as audio arrives, with interim and final results. The is_final indicator distinguishes text that may be revised from a finalized segment. The speech_final indicator relates to endpointing of a spoken segment. Keep these signals distinct from merely receiving a non-empty transcript, and retain the final segments in their audio order when preparing the completed record.

Nova-3 is general ASR rather than the separate Flux conversational model with model-native turn detection. Nova marks transcript finality and speech endpoints; Flux supplies its own conversational turn-state events. Choose the transcription workflow around whether the audio is already recorded or is arriving live, while using the actual final output as the artifact to retain.

Keyterms for vocabulary that matters

Nova-3’s documented Keyterm Prompting supports monolingual and multilingual transcription and adapts recognition toward important terms without retraining the model. Supply relevant names, product terms, acronyms or multi-word phrases that may actually be spoken. The guide limits the combined keyterm input to 500 tokens and recommends focusing on the most important vocabulary rather than adding common words indiscriminately.

The producer’s syntax uses plain terms, with a separate keyterm entry for each term. A multi-word phrase stays one phrase. Legacy keyword weights or intensifiers, and joining separate terms with commas, semicolons or line breaks, are not equivalent: the interface can accept them as one literal value while providing no intended boost. Use the correct capitalization for names, then check the transcript because audio context and smart formatting can still change the final written form.

Model-family boundaries and useful results

The general Nova-3 profile is distinct from the separately named Medical and Pharma options. Vocabulary prompting does not turn it into another edition, and specialized models’ language coverage or behavior should not be assumed for the general model. Additional transcript-processing features are choices in the producer’s speech workflow, not evidence that recognition has produced a summary, translation or other artifact that was never requested.

A useful Nova-3 deliverable can be a searchable meeting transcript, caption-ready wording with word timings, or a record of a multilingual conversation. Retain the audio identifier, selected language workflow, final text and any useful confidence or timing fields. Check unclear names, numbers, overlapping voices and language changes against the original recording, and keep the actual recognized speech available for the next content-processing step.