NVIDIA Parakeet TDT 0.6B V3 ASR

Overview

Parakeet TDT 0.6B V3 is a multilingual automatic speech recognition model from NVIDIA. It uses a FastConformer encoder plus a Token-and-Duration Transducer decoder to map audio to text. It auto-detects the spoken language and returns punctuated, capitalized transcripts. This helps you turn long recordings into searchable text, captions, and logs.

What you can build

• Meeting and interview transcripts with readable punctuation and casing
• Subtitle drafts for videos and podcasts
• Voice note transcription for knowledge bases and task systems
• Search and indexing for large audio archives
• Call review and QA workflows that need fast first-pass transcripts

Inputs

• One or more audio files to transcribe. Use WAV or MP3 for best results. Other formats can be converted before transcription.
• Audio that is close to 16 kHz, single-channel speech. Strong compression or stereo mixes can reduce accuracy.
• An option to include timestamps. Enable it when you need word-level and segment-level timing.
• An optional batching control for multi-file jobs. Larger batches can increase throughput but need more memory.

Outputs

The model returns a transcription for each provided audio file as text. The text typically includes punctuation and capitalization. When you enable timestamps, the output also includes structured timing data. That timing data includes start and end times in seconds for segments and individual words.

Limitations

• Language support is limited to 25 European languages. It won’t transcribe languages outside that set.
• The model auto-detects language. You generally can’t force a specific language.
• It does speech-to-text only. It does not do speech-to-text translation.
• Very long recordings may require chunking or long-form settings. Accuracy can drift over long durations.
• It does not label speakers. Overlapping speech can reduce accuracy.
• Low-quality audio can cause errors. This includes clipped speech, strong background music, heavy noise, and low bitrates.

Safety & compliance

• Follow the CC BY 4.0 license terms when you use and redistribute outputs.
• Get consent before transcribing private conversations.
• Treat transcripts as sensitive data when audio includes personal information.
• Review outputs before using them for legal, medical, hiring, or other high-impact decisions.