When using Deepgram's API with the nova-2 model, users might occasionally encounter an issue where correct transcriptions are initially received, but after approximately 8 minutes of streaming, the transcriptions return empty results with confidence dropping to zero. Despite the lack of transcription, charges still apply, leading to concerns and confusion.
If you're sending small, sporadic audio chunks (such as 2-10 frames per minute), Deepgram may expect a more continuous audio input. Intermittent input might cause some models, including nova-2, to behave unexpectedly in returning results.
To address this issue, it is recommended to:
If issues persist, or behavior continues to be inconsistent, reaching out to Deepgram support is recommended. Visit our community for more help: Deepgram Discord Community.
For optimal use of Deepgram's API, especially with models like nova-2 that may have specific input handling characteristics, ensure your audio streams are continuous. Adapt APIs to make use of callback mechanisms for handling intermittent input scenarios.
Always ensure your implementation uses the most current SDK versions and follows best practices for network stability and audio quality.