Gabriel BibbóSource code
Speech enhancement · Backend platform · MLOps · 2026

A reproducible backend for testing whether enhancement helps ASR

The platform compares two transcription paths on the same pre-recorded audio: direct ASR and speech-enhance-then-ASR. It keeps the complete job state, audio artifacts, transcripts and provider responses inspectable so enhancement can be evaluated as a system rather than as an isolated DSP step.

2 pathsraw vs enhance-and-transcribe
4 presetsbypass through denoise + dereverb
8 servicesbackend Compose stack
CI + telemetrytests, metrics, logs and traces

▶️ Live demo — try it now

https://asr-rp5.tail072b8f.ts.net/demo/

When the browser asks, sign in with:

Username
recruiter
Password
asr-demo-2026

What the platform does

Controlled comparison

The same input can go directly to transcription or through an enhancement preset first, making the effect of denoising and dereverberation explicit and repeatable.

Persistent jobs

Job state lives in PostgreSQL while raw audio, enhanced audio, transcripts and provider payloads use deterministic object-storage paths in MinIO.

Asynchronous processing

FastAPI handles the HTTP surface and Celery workers execute the processing path through Redis, separating request handling from longer-running audio work.

Provider abstraction

A deterministic fake provider supports local development and CI without credentials, while live pre-recorded ASR providers can be enabled behind explicit opt-in gating.

System architecture

API
FastAPI exposes health, readiness, transcription, enhancement-and-transcription, job-result and metrics endpoints.
Worker
Celery runs asynchronous audio and transcription jobs.
State
PostgreSQL stores job state, Redis acts as the worker broker and MinIO stores audio and transcript artifacts.
Observability
JSON structured logs, Prometheus metrics, OpenTelemetry traces and a provisioned Grafana dashboard make the backend inspectable during execution.
Deployment
Docker Compose provides a reproducible single-host deployment path for the backend stack.

Enhancement pipeline

The current DSP path exposes four deterministic presets: bypass, light clean, denoise, and denoise plus dereverberation. This keeps the processing choices explicit and versionable while the surrounding platform handles persistence, orchestration and inspection.

ASR Enhancement Platform
The project combines audio processing with the backend infrastructure needed to run and inspect repeatable transcription jobs.

Engineering scope

The current MVP targets pre-recorded audio and prioritises reproducibility, observability and a complete backend vertical slice. It is not a production-hardened service and does not claim high-availability infrastructure, enterprise authentication or production-scale orchestration.

Repository
github.com/gbibbo/asr_enhancement
FastAPICeleryDockerPostgreSQLRedisMinIOPrometheusOpenTelemetryGrafana