Truncated recognitions leads to repetitions in questions

[Lovenoreus]

At 01:40 I say the address "Yes this happened on Main street one" and the bot hears this in the partial recognition as "This Main Street one. yes the seven on" and I don't know if this is the normal quality of Azure voice recognition or a product of this repo

container_call_2.mp4

container_log_2_splitting_of_messages.txt

Transcript.written.before.the.call.that.i.will.follow.t.txt

[clemlesne]

If I understand well, some voice samples are truncated by the bot, like in this example:

• Said: Yes... this happened on main street... one
• Stored: This Main Street one. yes the seven on call.

Indeed, the recognition is bad in that case. Without debugging logs, we cannot distinguish the partial and complete recognitions.

Possible root causes:

• I can hear echo while you recorded your voice. Was your call also on speaker? If yes, the echo cancellation algorithm applies noise reduction, which lowers the voice quality (this is a tradeoff with OSS software while performing in real-time).
• AI Speech speech to text temporary degraded mode
• You made multiple pauses in the sentence, and the TTS pause timeout is 250 ms by default, the timeout could have been triggered before the sentence end, resulting in truncated text

[Lovenoreus]

I did not say a larger sentence that was being truncated into "Yes... this happened on main street... one". If that is the case then it is fetching text from some earlier sentence and merging them.

I did attach the log file and I will also show a small relevant part of it bellow, it is with enabled debugging logs like so:
monitoring:
logging:
app_level: DEBUG
sys_level: INFO

2025-01-21T15:15:02.055235Z [info ] Feature slow_llm_for_chat not found, using default: True call.channel=voice call.id=18167bc5-df67-454a-b32b-e6eeef6f626c call.message=This Main Street one. yes the seven on call.phone_number=+4570726778
2025-01-21T15:15:02.056843Z [info ] Using 20/20 messages (6106 tokens) as context call.channel=voice call.id=18167bc5-df67-454a-b32b-e6eeef6f626c call.message=This Main Street one. yes the seven on call.phone_number=+4570726778
/app/app/helpers/call_llm.py:729: RuntimeWarning: coroutine 'SttClient._report_complete_latency' was never awaited
silence_task = None
RuntimeWarning: Enable tracemalloc to get the object allocation traceback
2025-01-21T15:15:02.514494Z [info ] Stoping TTS after 250 ms call.id=18167bc5-df67-454a-b32b-e6eeef6f626c call.phone_number=+4570726778
2025-01-21T15:15:02.535436Z [info ] Stoping TTS after 250 ms call.id=812d0408-b07d-4e39-9a64-536dab9b2ea1 call.phone_number=+4570726774
2025-01-21T15:15:02.556733Z [info ] Stoping TTS after 250 ms call.id=812d0408-b07d-4e39-9a64-536dab9b2ea1 call.phone_number=+4570726774
2025-01-21T15:15:02.595065Z [info ] Playing TTS: Thank you for the information. call.channel=voice call.id=18167bc5-df67-454a-b32b-e6eeef6f626c call.message=This Main Street one. yes the seven on call.phone_number=+4570726778
2025-01-21T15:15:02.719771Z [info ] Playing TTS: I will update the claim with the location as Main Street. call.channel=voice call.id=18167bc5-df67-454a-b32b-e6eeef6f626c call.message=This Main Street one. yes the seven on call.phone_number=+4570726778
2025-01-21T15:15:02.815086Z [info ] Stoping TTS after 250 ms call.id=af0f6372-0532-44fd-89cc-66235aaa49a3 call.phone_number=+4570726778
2025-01-21T15:15:02.832017Z [info ] Playing TTS: Is there anything else you would like to add about the incident? call.channel=voice call.id=18167bc5-df67-454a-b32b-e6eeef6f626c call.message=This Main Street one. yes the seven on call.phone_number=+4570726778
2025-01-21T15:15:03.633905Z [info ] Voice stored: yes the seven on main street one call.id=18167bc5-df67-454a-b32b-e6eeef6f626c call.phone_number=+4570726778
2025-01-21T15:15:03.659107Z [info ] Enhancing LLM chat with 0 trainings call.channel=voice call.id=18167bc5-df67-454a-b32b-e6eeef6f626c call.message=yes the seven on main street one call.phone_number=+4570726778
2025-01-21T15:15:03.684664Z [info ] Feature slow_llm_for_chat not found, using default: True call.channel=voice call.id=18167bc5-df67-454a-b32b-e6eeef6f626c call.message=yes the seven on main street one call.phone_number=+4570726778
2025-01-21T15:15:03.685712Z [info ] Using 20/20 messages (6030 tokens) as context call.channel=voice call.id=18167bc5-df67-454a-b32b-e6eeef6f626c call.message=yes the seven on main street one call.phone_number=+4570726778
2025-01-21T15:15:05.293189Z [info ] Playing TTS: I am updating the claim with the location as Main Street 7. call.channel=voice call.id=18167bc5-df67-454a-b32b-e6eeef6f626c call.message=yes the seven on main street one call.phone_number=+4570726778 tool.args={"updates": [{"field": "incident_location", "value": "Main Street 7"}], "customer_response": "I am updating the claim with the location as Main Street 7."} tool.name=updated_claim

[clemlesne]

I don't see any issues in the log file you sent

[clemlesne]

Commit a6a3958 should fix the ability to cut answers and properly load the next. It may mitigate your issue. Publshed as of v17.1.0.