Repository navigation
v3.1.0
v3.1.0 (2026-09-24)
This release is published under the Apache-2.0 License.
Bug Fixes
-
Beam search test result (
a0480fb) -
Device scope (
9379765) -
Github actions (
aedec2f) -
Github actions (
284c444) -
Install tpu (
5f365b0) -
Kaggle lm train define accelerator (
57ede0e) -
Kenlm training on kaggle (
a3b3b85) -
Lm training (
802394b) -
Memory for multi-head attention layer (
7076e72) -
Refactor lm (
f057eff) -
Split lm dataset, update kaggle (
07183ef) -
Step per epoch (
409d709) -
Streaming should have states for conv and mha layers (
74b5313) -
Text path (
8885de9) -
Tf 2.19 to enable tf.lite flex delegate ops (
e2228bf) -
Tflite convert and run (
ab6f29c) -
Tflite convert metadata (
586c36f) -
Tflite converter (
1161b80) -
Tmp dir for kenlm (
4bc7a41) -
Train lm loss (
05476c5) -
Train lm model upload checkpoint (
9a8dbdb) -
Unittests for tflite (
fbe4c91) -
Update lm dataset and training (
6aa7088) -
Update lm dataset and training (
78b3544) -
Update streaming with kv-cache (
cf90972) -
Update test with beam search lm (
29c680e) -
Update tests (
9a27ccd)
Chores
Documentation
- Lstm lm with training (
70d2472)
Features
-
Add ASRInference (
c5f296b) -
Add beam search with lm separately from beam search only (
5451cf6) -
Add examples on ASRInferences and update docs (
50a9fab) -
Add norm lm dataset (
fb7508f) -
ALSD transducer beam search, ruff+uv refactor (
0886de7) -
Ilme,lodr for beam search transducer (
74e730e) -
Kenlm ngram load via arpa (
b9db071) -
Lstm lm with training (
c6739c5) -
Lstm lm with training (
561d4d9) -
Tpu support for train lm (
b8f66ec)
Detailed Changes: v3.0.0...v3.1.0