Splicr
A durable long-document speech pipeline designed around resumability, natural chunk boundaries, and provider independence.
- Role
- Product architecture and full-stack implementation
- Context
- Built to make long-form document-to-audio work observable and recoverable instead of treating it as one fragile request.
PythonFastAPISQLiteBackground processingDockerCI
Problem
Long documents exceed ordinary speech request limits and make failures expensive when progress, chunk boundaries, and provider state are not persisted.
Ownership and context
Product architecture and full-stack implementation. Built to make long-form document-to-audio work observable and recoverable instead of treating it as one fragile request.
Implementation
- Separated document import, natural-boundary chunk planning, speech generation, and WAV assembly into explicit stages.
- Persisted checkpoints so interrupted work can resume or retry from a known state.
- Defined provider-neutral adapters so orchestration is not coupled to one speech API.
Recoverable long-form generation
- Durable checkpoints limit the cost of interrupted or partially failed jobs.
- The browser studio exposes document, chunk, and job state without hiding the underlying workflow.