Spaces:
Running on Zero
Running on Zero
Commit History
24 days ago
-
Serve the app this repository maintains Running
Verified signatureThis Space carried its own copy of the OWSM v4 demo, pinned to espnet v.202412 and drifting from the one under review in the ESPnet repository. It now runs the same app.py and requirements.txt as espnet/owsm-v4, both published from egs2/owsm_v4/s2t1/demo. The card is the one file that differs: it says this is the same demo at an older address, kept because the address is linked from elsewhere. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>a03f0d8 -
Load in bfloat16, keep long-form end detection, drop the config patch
Verified signatureThree changes, now that the pinned espnet is current. The block that rewrote gradient_checkpoint_layers out of the cached config.yaml is gone. That key became a real argument of the E-Branchformer encoder and the transformer decoder in January 2025; it was only ever a workaround for v.202412, and it wrote inside the model cache to do it. All three models load with dtype=bfloat16. A 1.02B model in fp32 is about 4.1 GB of weights, and three of them are resident. Note that the log-Mel frontend is cast along with the rest of the model, so this needs listening to, not only timing. Long-form decoding no longer sets maxlenratio = -300, which disabled the beam search's end detection. 0.0 is the same maximum output length and stops once the search has converged: on LibriSpeech test-clean (87 chapters, OWSM v4 base, beam 5, one V100) the WER was identical and RTF went from 0.142 to 0.102. The page's own description said to use 1.0, which is not the default and not what the recipe recommends; …81ec4a0 -
Ask for the Python the pinned espnet requires
Verified signatureespnet declares requires-python >=3.12,<3.14. With no python_version here the Space builds on whatever its image defaults to, and the install fails before it reaches the model. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>d93f6fe -
Pin espnet to a current master, and torch/numpy to what it needs
Verified signaturev.202412 predates the decoding work the demo should be showing: batch invariance (#6633/#6660), the end-detection cache (#6667), and the CTC fixes merged on 2026-10-09 (#6856/#6857/#6868). #6856 in particular is this Space's own CTC path: the window could not cross a silence longer than itself, which inserts words around long pauses. That espnet needs torch>=2.11.0,<2.15 with torchaudio>=2.11.0,<2.12, and numpy>=2.0.0, so 2.10.0 and numpy 1.23 no longer resolve. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>cf011d1
-
Update requirements.txt
Verified signatureUpdate torch version from 2.4 to 2.10d26c3f3
-
Update app.py
Verified signature0dc1353 -
Fixed config path
Verified signature50ee433
-
Update app.py
Verified signaturee33a5ab -
Update requirements.txt
Verified signature8870640 -
Update requirements.txt
Verified signaturea1261c2 -
Update app.py
Verified signature3dfdefc -
Update requirements.txt
Verified signaturee323307 -
Update requirements.txt
Verified signature91182c1 -
Update app.py
Verified signatureae97e77 -
Update app.py
Verified signatureUpdated description from v3.1 to v4cd5c512 -
Update app.py
Verified signature00de4b3 -
Update app.py
Verified signatured641d01 -
Update app.py
Verified signature5dcded9 -
Upload app.py
Verified signature1f20047 -
Create requirements.txt
Verified signatureca7a790 -
initial commit
Verified signature788588d