NVIDIA: Parakeet TDT 0.6B v3

NVIDIA: Parakeet TDT 0.6B v3 is a speech-to-text model from NVIDIA. It is a 600 million parameter model built on the FastConformer-TDT architecture, and it was trained on the Granary dataset, which contains more than 670,000 hours of audio. The model is positioned as a multilingual transcription system, and it supports automatic language detection across the languages it covers. If you are already working within the NVIDIA ecosystem, or you simply need a compact, focused transcription model, this is the kind of model you would evaluate on the strength of its architecture and training data rather than on general reasoning ability.

The model is single purpose. Its only input modality is audio, and its only output modality is transcription. It does not accept text prompts in the usual chat sense, and it does not generate conversational replies. Because it is a dedicated transcription model rather than a general purpose language model, there is no context window measured in tokens and no maximum output length in the way a chat model would report one; the relevant limits are about audio length and format rather than prompt size. There are no supported request parameters listed for this model on api14, so you should expect to send a straightforward transcription request without tuning options.

The use cases follow directly from that design. It fits batch or on demand transcription of recorded audio, such as meeting recordings, interviews, voice notes or call recordings. Because it performs automatic language detection across multiple languages, it can be a reasonable choice when you do not know the language of the audio in advance and do not want to route each file manually. It also suits pipelines that need a compact, dedicated speech-to-text step rather than a large general model doing transcription as a side effect of a broader prompt.

You call it through api14's OpenAI-compatible API using the model id "nvidia/parakeet-tdt-0.6b-v3". Point your existing OpenAI client at the api14 base URL, send your audio as the input, and read the transcription from the response. Since no request parameters are listed as supported for this model, keep your request minimal and do not rely on temperature, top_p, max_tokens or similar sampling controls having any effect. Treat it as a transcription endpoint that happens to share the OpenAI request shape.

One practical limit to keep in mind is that this model only does transcription. If your task requires summarising the transcript, answering questions about it, translating it into another language, or any other text generation, you will need a separate text model for that step, or a different model that handles both audio input and generative output. Similarly, if you need configurable decoding behaviour or fine grained control over the output, check the live parameter list before building around it, because none are listed here. For plain speech-to-text with language detection, it is a direct fit.