parakeet-unified vs nemotron-speech-streaming?
#7
by xdevfaheem - opened
how are they different? I'm only focusing english model, which should i go with?
Hi @xdevfaheem , sorry for the delayed answer.
The main differences are:
- The Unified model works well in offline and streaming decoding.
- The Unified model outperforms Nemotron down to 240ms latency (according to the Open ASR leaderboard). But if you need 160ms or 80ms, it is better to use the Nemotron model.
- The Unified model does not support cache-aware decoding yet (only chunked inference with left context recomputation). You can observe 100-300% RTFx degradation at low latencies for streaming decoding compared to Nemotron.
What latency are you targeting for English?