Self Hosted: Transcription is not working

[Quentinchampenois]

Hello,

I am opening this discussion to have help or clues about the issue I am facing on my self-hosted app. It appears that as a user on the portal, I never get the transcription.
As an administrator, I can start a discussion from the dashboard and assistant answers.

I do have audio chunks into Directus and audio files into my S3 storage.

Did you ever encountered this behaviour before ?

Echo version: 1.3.0

Here is the entire Helm chart : https://github.com/OpenSourcePolitics/echo/pull/1/files

Thanks in advance

[Quentinchampenois]

I configured the DEBUG_MODE on my worker and now I found this error :

INFO:celery.worker.strategy:Task dembrane.tasks.task_finish_conversation_hook[45019b28-323d-4417-876a-c1e01783a97c] received
[2025-04-30 20:56:50,454: INFO/ForkPoolWorker-4] dembrane.tasks.task_finish_conversation_hook[45019b28-323d-4417-876a-c1e01783a97c]: transcript is empty for conversation: 3119a399-294f-41f4-8135-84ca48d0ad89. so not generating summary
INFO:celery.app.trace:Task dembrane.tasks.task_finish_conversation_hook[45019b28-323d-4417-876a-c1e01783a97c] succeeded in 0.10175113612785935s: None
INFO:celery.worker.strategy:Task dembrane.tasks.task_process_conversation_chunk[1557657c-1f5d-46a2-8f91-40d067d9d484] received
[2025-05-02 06:50:09,981: ERROR/ForkPoolWorker-4] dembrane.tasks.task_process_conversation_chunk[1557657c-1f5d-46a2-8f91-40d067d9d484]: Error: (psycopg.OperationalError) consuming input failed: SSL SYSCALL error: EOF detected
[SQL: SELECT conversation_chunk.id AS conversation_chunk_id, conversation_chunk.created_at AS conversation_chunk_created_at, conversation_chunk.updated_at AS conversation_chunk_updated_at, conversation_chunk.conversation_id AS conversation_chunk_conversation_id, conversation_chunk.path AS conversation_chunk_path, conversation_chunk.timestamp AS conversation_chunk_timestamp, conversation_chunk.transcript AS conversation_chunk_transcript
FROM conversation_chunk
WHERE conversation_chunk.id = %(pk_1)s::UUID]
[parameters: {'pk_1': 'b133e9ed-8a9d-44e1-bc9e-ce72aeb17baf'}]
(Background on this error at: https://sqlalche.me/e/20/e3q8)
INFO:celery.worker.strategy:Task dembrane.tasks.task_process_conversation_chunk[1557657c-1f5d-46a2-8f91-40d067d9d484] received
INFO:celery.app.trace:Task dembrane.tasks.task_process_conversation_chunk[1557657c-1f5d-46a2-8f91-40d067d9d484] retry: Retry in 180s: OperationalError('(psycopg.OperationalError) consuming input failed: SSL SYSCALL error: EOF detected')
INFO:celery.worker.strategy:Task dembrane.tasks.task_process_conversation_chunk[104dc4a2-4cbe-46bc-a628-c5fa8f579967] received
INFO:celery.app.trace:Task dembrane.tasks.task_process_conversation_chunk[104dc4a2-4cbe-46bc-a628-c5fa8f579967] succeeded in 0.030525009147822857s: <AsyncResult: d58d54fd-62db-4fa0-98ca-591fa157452f>

[Quentinchampenois]

Error SSL SYSCALL error: EOF detected' seems to be solved with engine = create_engine(DATABASE_URL, pool_pre_ping=True)
cf: https://docs.sqlalchemy.org/en/14/core/engines.html#sqlalchemy.create_engine.params.pool_pre_ping

[ussaama]

@Quentinchampenois sorry for the late reply, the error likely occured from a stale or dropped PostgreSQL connection which can happen for various reasons like network interruptions, database restarts, idle connection timeouts or proxies etc silently closing connections. pool_pre_ping=True seems to ensure SQLAlchemy checks if a connection is still valid before reusing it. Has this resolved your issue or are you still running into problems?

[Quentinchampenois]

Hello @ussaama thanks for the reply, indeed I didn't fixed it yet.
In fact, when I am recording :

• Audio chunks are sent to Directus and saved
• In the S3, I have these records
  But I didn't found where the chunk is sent for transcription to the third-party AI service. In my logs I just found messages like "Transcription empty", but I don't get what is in charge of transcription in the server side

[ArindamRoy23]

Hey,

Please look at the 'server/dembrane/transcribe.py' and the function 'transcribe_audio_litellm'. It uses the litellm package to access whiper. This can be configured to any type of standard whisper deployment (AZ/OpenAI).