On May 1, 2025, Nvidia introduced Parakeet-TDT-0.6B-v2, a language model designed for automatic speech recognition (ASR). The developers claim that the new open-source LLM is capable of transcribing an hour of audio in just one second, which is equivalent to the results of proprietary analogues such as GPT-4o and ElevenLabs Scribe. Parakeet-TDT-0.6B-v2 is distributed under a free CC-BY-4.0 license, allowing even commercial projects to integrate it into their applications.
After its release on the Hugging Face platform, the Parakeet-TDT-0.6B-v2 language model became the leader of the Open ASR Leaderboard, which includes the best open-source LLMs. According to the Word Error Rate (WER) metric for evaluating the average number of errors, the new neural network achieved a 6.05% accuracy loss, which is significantly higher than the GPT-4o-transcribe model, which has a 2.46% accuracy loss, and ElevenLabs Scribe with 3.3%. It is noteworthy that Microsoft’s Phi-4 language model ranks second in the rating, but it also does not support Russian, just like Parakeet-TDT-0.6B-v2. In other metrics, such as LibriSpeech WER for evaluating clear speech recognition, the neural network has 1.69% accuracy, and in the SNR 5 benchmark for evaluating noisy text recognition, the accuracy loss was 8.39%. In telephone call processing tasks, where audio is compressed via μ-law, accuracy losses are also minimal, at only 4.1%.
The neural network is capable of ignoring external noise, can place punctuation and accurate time stamps for each word, and also supports the transcription of songs and telephone conversations, which is especially useful when used in business applications. It is also worth noting that Parakeet-TDT-0.6B-v2 supports .wav and .flac audio formats with a frequency of 16 kHz. Although the model is optimized to run on Nvidia GPUs (A100, H100, T4, V100), thanks to its 600 million parameters, Parakeet-TDT-0.6B-v2 can be run on low-end computers with 2 GB of RAM and even smartphones.
The architecture of the Nvidia Parakeet-TDT-0.6B-v2 model includes FastConformer and TDT. Fast Conformer is a modified Conformer architecture that significantly speeds up speech recognition by increasing the downsampling parameter by a factor of 8, which is achieved through the use of lightweight software elements and a combined attention mechanism with improved context understanding. TDT is a special decoder that predicts words, sounds, and their duration. The unique feature of TDT is that it focuses only on important audio elements, without wasting tokens on unnecessary speech segments such as letter elongations and pauses. This reduces the consumption of computing resources and speeds up speech processing without losing accuracy. Thanks to these technologies, the neural network is capable of processing audio recordings 3,386 times faster than manual transcription of a 128-byte audio packet. Python scripts and the NeMo framework are available for integrating the new language model into applications.
Parakeet-TDT-0.6B-v2 was trained on 128 A100 GPUs for 10,000 hours using synthetic data and 120,000 hours of real human speech from YouTube videos and phone conversations. Part of the Granary dataset used to train the model is not yet available to users, but Nvidia has stated that it will open access to it after the Interspeech 2025 conference. The company also emphasized that no personal user data was used in the training, and the documentation includes a description of the data collection and privacy assessment methods.
Conclusions
Independent developers have already hailed Parakeet-TDT-0.6B-v2 as a breakthrough in the field of neural networks for speech processing and even the entire open source community. Thanks to the new neural network from Nvidia, companies and enthusiasts around the world will be able to use compact and efficient artificial intelligence to create transcription services, voice assistants, and subtitle generators, completely free of charge. The release of Parakeet-TDT-0.6B-v2 demonstrates Nvidia’s strengthening position in the market for developing advanced LLM models, gradually challenging AI giants such as OpenAI and Google.
