Parakeet

Parakeet is NVIDIA's speech recognition model family. The linked English model, parakeet-tdt-0.6b-v2, is for developers building transcription services, subtitles and voice applications with a model they can run locally.

• Readable transcripts: Convert speech to text with automatic punctuation and capitalization, including support for spoken numbers and song lyrics.
• Audio timestamps: Obtain word, character and segment timings to connect transcript text with moments in a recording.
• Long recordings: Transcribe audio segments of up to 24 minutes in a single pass using the FastConformer-TDT architecture.
• Fine-tuning: Use the pretrained checkpoint directly or adapt it to another dataset through NVIDIA NeMo.

Local inference runs through NeMo and PyTorch on Linux and accepts mono WAV or FLAC audio. NVIDIA GPUs are recommended for inference; NeMo also supports CPU execution. Training requires an NVIDIA GPU with CUDA. The model needs at least 2 GB of RAM to load, with more memory needed for longer audio.

The weights use CC BY 4.0, permitting commercial and noncommercial use with attribution. The NeMo Speech framework is open source under Apache 2.0. NVIDIA also offers a hosted API that receives your audio and requires an API key and internet access, without a local GPU.

These capabilities apply to the linked English checkpoint. Accuracy varies with accents, noise and speech context; isolated words and incomplete sentences can be harder to recognize.