v1.0: ltd26 study checkpoint

.gitattributes

@@ -33,3 +33,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
+model.nemo filter=lfs diff=lfs merge=lfs -text

README.md

@@ -0,0 +1,53 @@
+---
+license: cc-by-4.0
+language:
+- lt
+library_name: nemo
+pipeline_tag: automatic-speech-recognition
+base_model: nvidia/parakeet-tdt-0.6b-v3
+datasets:
+- Digisensus/lithuanian-dialect-speech-liepa-3-100h-punctuated
+- Digisensus/lithuanian-phone-speech-liepa-3-429h-punctuated
+- Digisensus/lithuanian-dialect-asr-eval
+tags:
+- speech
+- lithuanian
+- dialects
+- parakeet
+- tdt
+---
+
+# parakeet-tdt-0.6b-lt-study-s-d80-std
+
+Research checkpoint from the study *Lithuanian speech recognition: effects of dialect training and
+transcript spelling*. Fine-tuned from `nvidia/parakeet-tdt-0.6b-v3` on LIEPA-3 telephone speech (420.17 h) + LIEPA-3 dialect speech (82.30 h), dialect clips in standard spelling.
+Output is lowercase spoken form without punctuation. This is a study model, not a product model.
+
+## WER (%)
+
+| Test | WER |
+|---|---|
+| Dialect test, dialect-spelling reference | 37.52 |
+| Dialect test, standard-spelling reference | 27.06 |
+| Dialect test, either spelling accepted | 26.06 |
+| LIEPA-3 telephone test | 10.65 |
+| FLEURS lt test | 26.87 |
+| Common Voice 19 lt test | 27.75 |
+
+## Use
+
+```python
+import nemo.collections.asr as nemo_asr
+model = nemo_asr.models.ASRModel.from_pretrained("Digisensus/parakeet-tdt-0.6b-lt-study-s-d80-std")
+print(model.transcribe(["audio.wav"])[0].text)
+```
+
+## Provenance
+
+| | |
+|---|---|
+| Run | `ltd26/S+D80.std/s1/a1` |
+| Recipe | `S+D80.std` (`recipes/S+D80.std.json` in the code repo) |
+| Splits | `splits-v1` |
+| Training | 10,000 updates, 192 clips per update, last checkpoint; full spec in `config.json` |
+| Scorer | see `registry/scores.csv` in the code repo |

config.json

@@ -0,0 +1,69 @@
+{
+  "engine": "parakeet",
+  "run_id": "ltd26/S+D80.std/s1/a1",
+  "model": "nvidia/parakeet-tdt-0.6b-v3",
+  "head": "tdt",
+  "parakeet": {
+    "head": "tdt",
+    "head_lr_multiplier": 10.0,
+    "lowercase": true,
+    "strip_punctuation": true,
+    "normalize_quotes": true,
+    "drop_unk_chars": true,
+    "min_seconds": 0.5,
+    "max_seconds": 30.0,
+    "batch_size": 48,
+    "max_batch_seconds": 200.0,
+    "max_joint_gb": 256.0,
+    "max_tokens_per_sec": 12.0,
+    "num_workers": 12,
+    "seed": 1,
+    "num_epochs": 1000,
+    "max_steps": 10000,
+    "learning_rate": 0.0001,
+    "min_lr": 1e-06,
+    "warmup_steps": 500,
+    "warmup_ratio": 0.05,
+    "weight_decay": 0.001,
+    "gradient_accumulation_steps": 4,
+    "max_grad_norm": 1.0,
+    "freeze_encoder": false,
+    "spec_augment": true,
+    "loss_impl": "auto",
+    "joint_stream": true,
+    "fused_batch_size": 16,
+    "joint_budget_gb": 0.0,
+    "eval_every_steps": 1000,
+    "eval_max_samples": 2000,
+    "eval_batch_size": 32,
+    "logging_steps": 25,
+    "keep_ckpt_dirs": 3,
+    "final_model": "last",
+    "early_stopping_patience": 0,
+    "early_stopping_threshold": 0.0
+  },
+  "loss_impl": "numba",
+  "precision": "bf16",
+  "init_from": "",
+  "torch": "2.11.0+cu128",
+  "nemo": "2.7.3",
+  "step": 10000,
+  "epoch": 1000,
+  "final_from": "last",
+  "final_step": 10000,
+  "metrics": {
+    "wer": 15.07555115184543,
+    "cer": 4.63790788254798,
+    "wer_norm": 15.07555115184543,
+    "subs": 12.200478903476178,
+    "ins": 1.1724878209891834,
+    "dels": 1.7025844273800677,
+    "n_utts": 7380.0,
+    "precision": 86.55576584653187,
+    "recall": 86.09693666914376,
+    "f1": 86.32574158242886
+  },
+  "train_hours": 502.48,
+  "train_n": 499495,
+  "eval_n": 7380
+}

model.nemo

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:cfd85aea78e42b62cb9dc21968789ef999514d82bb004f4625ba8efb8761d569
+size 2509342720