Release Notes

Release 1.6.0

Key Features and Enhancements

• Added new models with Silero VAD and Sortformer diarizer support:

  • Parakeet CTC 0.6b Vietnamese (vi-VN)

  • Parakeet CTC 0.6b Mandarin (zh-CN)

  • Parakeet CTC 0.6b Spanish (es-US)

• Updated the Parakeet CTC 0.6b English (En-US) model with Silero VAD and Sortformer diarizer support.

• Updated the Parakeet CTC 1.1b English (En-US) model with FP8 profiles.

• Added support for Realtime API via Websockets for streaming ASR applications.

Known Issues

• Sortformer speaker diarization can occasionally produce incorrect speaker tags (like overlapping speech) in low latency mode.

• Profiles with vad=silero use Silero VAD for endpointing, which can result in the following issues:

  • Reduce ASR transcription accuracy.

  • Increase latency in streaming-throughput mode

  • Lower throughput in offline mode.

• The Parakeet CTC 0.6b Mandarin (zh-CN) model can occasionally insert double punctuation marks.

• The Parakeet CTC 0.6b Vietnamese (vi-VN) model can occasionally generate inaccurate timestamps.

Release 1.5.0

Key Features and Enhancements

• Added the Parakeet TDT 0.6b v2 English ASR model for accurate offline transcription with punctuation, capitalization, and word timestamp support.

• Updated the Parakeet 1.1b CTC English (en-US) NIM to support Speaker Diarization.

• Added language detection support for the Whisper Large v3 and Parakeet 1.1b RNNT Multilingual models.

• Added a profanity filter for the Whisper Large v3 and Parakeet 1.1b RNNT Multilingual models.

• Improved throughput for the Parakeet 1.1b RNNT Multilingual model.

Known Issues

• The Whisper Large v3 NIM does not support NVIDIA B200 GPUs.

• The Whisper Large v3 NIM uses OpenAI’s Whisper model, which may result in hallucinations, inaccuracies, and incorrect inverse text normalizations in some cases.

• The profanity filter may not work for non-English languages in some cases.

• The Parakeet 0.6b TDT model uses chunked inference and can sometimes produce incorrect transcripts compared to the NeMo model. Refer to Model Accuracy for details.

• Canary models may produce incorrect translations for certain languages such as Spanish (es-ES) and Korean (ko-KR).

• ASR HTTP API (Whisper) accepts only file, model, and language parameters in the request. All other parameters are ignored.

• ASR gRPC API (Whisper)

  • Supports only the offline Recognize API.

  • Does not support customization parameters (e.g. profanity_filter, enable_word_time_offsets, enable_automatic_punctuation, verbatim_transcripts) in the RecognitionConfig message.

• SSL/TLS encryption does not support TTS APIs. MTLS encryption is not supported in this release.