mega-asr-bench-fixed

Running

App Files Files Community

mega-asr-bench-fixed

Commit History

Jul 23

23:53
Update mega-asr.js

m2iceman

Verified signature
b13ee14

May 22

05:21
Clean up debug diagnostics now that WebGPU works end-to-end

Reza2kn

Verified signature
2cf4acc
05:15
Use INT4 encoder (MatMulNBits — WebGPU-supported); ORT 1.23 keeps webgpu fp16 fix for decoders

Reza2kn

Verified signature
b8c2d24
05:13
Per-step diagnostics: pinpoint which ORT call crashes

Reza2kn

Verified signature
7be4ffd
05:11
Surface full error info on transcribe failure

Reza2kn

Verified signature
50bb779
05:08
Upgrade onnxruntime-web 1.20 → 1.23 (webgpu bundle); revert WASM forcing — 1.23 fixes fp16 transformer NaNs

Reza2kn

Verified signature
e6b3306
05:03
Force decoders to WASM (WebGPU fp16 in ort-web 1.20 returns all-NaN on this transformer)

Reza2kn

Verified signature
d62ebc2
04:56
Diagnostics: logit type/len/health/first-8/top-5 with Number() coercion

Reza2kn

Verified signature
29d0bb3
04:55
Await tensor.getData() for WebGPU outputs (audio_embeds + logits) so data is actually copied back to CPU

Reza2kn

Verified signature
1320aec
04:53
fp16: use canonical u16 bit-pattern viewed as Float16Array; diagnostic top-5 dump

Reza2kn

Verified signature
ed37d13
01:58
Encoder: prefer static INT8 (QLinearConv); INT4 fallback. Recovers 92.7%-class quality in browser

Reza2kn

Verified signature
9e54b79
01:45
Float16Array for fp16 tensors; per-session WebGPU fallback (decoders stay on webgpu)

Reza2kn

Verified signature
af482b4
01:39
Encoder: try INT8 first, auto-fallback to INT4 if ConvInteger unsupported in browser

Reza2kn

Verified signature
a821180
01:38
Use ort.webgpu bundle + auto-fallback to wasm if webgpu init fails

Reza2kn

Verified signature
96f08e1
00:03
Bump cache key to invalidate RTN weights (GPTQ ship)

Reza2kn

Verified signature
36417e8

May 21

20:57
Use INT8 encoder + INT4 decoder (91.9% accuracy); force-English prompt default

Reza2kn

Verified signature
61dfe9b
20:46
Switch Space to static SDK: pure browser inference via onnxruntime-web

Reza2kn

Verified signature
a4c397e
20:46
Remove vendor/ (switching to static SDK)

Reza2kn

Verified signature
ffd2ab1
20:46
Remove requirements.txt (switching to static SDK)

Reza2kn

Verified signature
0dfc823
20:46
Remove app.py (switching to static SDK)

Reza2kn

Verified signature
da91cf8
20:39
Switch backend to INT4 ONNX models from Reza2kn/mega-asr-onnx

Reza2kn

Verified signature
888006f
20:27
Initial: Gradio demo + 8 VITW examples + WER scoring

Reza2kn

Verified signature
0c137e3
20:27
initial commit

Reza2kn

Verified signature
7feced8

Commit History

Jul 23

Update mega-asr.js

May 22

Clean up debug diagnostics now that WebGPU works end-to-end

Use INT4 encoder (MatMulNBits — WebGPU-supported); ORT 1.23 keeps webgpu fp16 fix for decoders

Per-step diagnostics: pinpoint which ORT call crashes

Surface full error info on transcribe failure

Upgrade onnxruntime-web 1.20 → 1.23 (webgpu bundle); revert WASM forcing — 1.23 fixes fp16 transformer NaNs

Force decoders to WASM (WebGPU fp16 in ort-web 1.20 returns all-NaN on this transformer)

Diagnostics: logit type/len/health/first-8/top-5 with Number() coercion

Await tensor.getData() for WebGPU outputs (audio_embeds + logits) so data is actually copied back to CPU

fp16: use canonical u16 bit-pattern viewed as Float16Array; diagnostic top-5 dump

Encoder: prefer static INT8 (QLinearConv); INT4 fallback. Recovers 92.7%-class quality in browser

Float16Array for fp16 tensors; per-session WebGPU fallback (decoders stay on webgpu)

Encoder: try INT8 first, auto-fallback to INT4 if ConvInteger unsupported in browser

Use ort.webgpu bundle + auto-fallback to wasm if webgpu init fails

Bump cache key to invalidate RTN weights (GPTQ ship)

May 21

Use INT8 encoder + INT4 decoder (91.9% accuracy); force-English prompt default

Switch Space to static SDK: pure browser inference via onnxruntime-web

Remove vendor/ (switching to static SDK)

Remove requirements.txt (switching to static SDK)

Remove app.py (switching to static SDK)

Switch backend to INT4 ONNX models from Reza2kn/mega-asr-onnx

Initial: Gradio demo + 8 VITW examples + WER scoring

initial commit