ai-spel/ai-spel-frontend
1
0
Fork 0
Code Issues Pull Requests Actions 14 Packages Projects Releases Wiki Activity

Use single-thread ONNX runtime for mobile speech
Build and publish frontend / publish (push) Successful in 50s
Details

Browse Source
This commit is contained in:
Ein Anderssono
2026-10-16 08:51:08 +02:00
parent aa832fd7cc
commit a4f89f2b82
9 changed files with 60 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files
docs/adr/0002-single-thread-onnx-for-mobile-tts.md
+13
View File
@@ -0,0 +1,13 @@
# ADR 0002: Run ONNX with one thread inside the Piper worker
## Decision
Configure ONNX Runtime Web to use one thread and stop backend initialization after 60 seconds. Piper already runs in a dedicated worker, so ONNX does not need to create another worker pool.
## Context
On iPad, the threaded ONNX WASM file downloaded fully, but the runtime did not report that initialization had completed. Nested worker startup and shared-memory setup are a likely compatibility boundary on mobile Safari. The available production assets currently include the threaded ONNX build.
## Consequences
This removes ONNX's additional worker pool and bounds initialization failures with a visible error. It may reduce inference throughput on desktop devices. CI and Kubernetes release verification must confirm that Piper still initializes and speaks on supported browsers; iPad acceptance remains a manual device check.