Description:
I am experiencing severe latency accumulation and "scrambled" audio transcripts when streaming audio to the google.genai V1 Beta API (gemini-2.0-flash-exp) over a persistent WebSocket connection.
Symptoms:
Latency Accumulation: On short utterances (<5s), response is fast. On longer utterances (20s+), the system does not seem to keep up with the real-time audio stream. At some point the system lags completely, I do not get transcript anymore and i have to wait around two minutes for Gemini to start responding again. Untill that the api is completely unresponsive and laggy.
Audio Scrambling/Garbage Transcripts: When this lag occurs, the server-side transcript eventually degrades into nonsense or scrambled text, suggesting that audio frames are being dropped or processed out of order due to the backlog.
Environment:
OS: macOS
Python: 3.9
SDK Version: google-genai (latest)
Model: gemini-2.0-flash-exp
Code Pattern: I am using a standard PyAudio producer / API consumer loop with asyncio.
Chunk Size: Tried 1024 (also tried 4096, same result).
Rate: 16kHz
Format: Int16 PCM
Method: session.send_realtime_input
Steps to Reproduce:
Open a client.aio.live.connect session.
Stream audio continuously from a microphone using send_realtime_input.
Speak continuously for 30-40 seconds.
Observe that the intermediate transcripts start lagging behind the spoken words, eventually becoming scrambled, and the model takes a long time to respond after silence.
Expected Behavior: The API should handle real-time 16kHz PCM streams without accumulating seconds of latency, or at least drop frames gracefully without corrupting the transcription stream.
Description:
I am experiencing severe latency accumulation and "scrambled" audio transcripts when streaming audio to the google.genai V1 Beta API (gemini-2.0-flash-exp) over a persistent WebSocket connection.
Symptoms:
Latency Accumulation: On short utterances (<5s), response is fast. On longer utterances (20s+), the system does not seem to keep up with the real-time audio stream. At some point the system lags completely, I do not get transcript anymore and i have to wait around two minutes for Gemini to start responding again. Untill that the api is completely unresponsive and laggy.
Audio Scrambling/Garbage Transcripts: When this lag occurs, the server-side transcript eventually degrades into nonsense or scrambled text, suggesting that audio frames are being dropped or processed out of order due to the backlog.
Environment:
OS: macOS
Python: 3.9
SDK Version: google-genai (latest)
Model: gemini-2.0-flash-exp
Code Pattern: I am using a standard PyAudio producer / API consumer loop with asyncio.
Chunk Size: Tried 1024 (also tried 4096, same result).
Rate: 16kHz
Format: Int16 PCM
Method: session.send_realtime_input
Steps to Reproduce:
Open a client.aio.live.connect session.
Stream audio continuously from a microphone using send_realtime_input.
Speak continuously for 30-40 seconds.
Observe that the intermediate transcripts start lagging behind the spoken words, eventually becoming scrambled, and the model takes a long time to respond after silence.
Expected Behavior: The API should handle real-time 16kHz PCM streams without accumulating seconds of latency, or at least drop frames gracefully without corrupting the transcription stream.