orukeet / transcribe-cpp /convert-orukeet.patch
NathanRoll's picture
Add validated transcribe.cpp Q8 export of Orukeet r3
b5f8f87 verified
Raw
History Blame Contribute Delete
1.64 kB
diff --git a/scripts/convert-parakeet.py b/scripts/convert-parakeet.py
index e980852..72bc7d9 100644
--- a/scripts/convert-parakeet.py
+++ b/scripts/convert-parakeet.py
@@ -121,6 +121,22 @@ V3_LANGUAGES = [
]
VARIANT_PROFILES: dict[str, dict] = {
+ "orukeet": {
+ "variant": "tdt-0.6b-orukeet",
+ "display_name": "Orukeet",
+ "version": "v0.1.0",
+ "size_label": "0.6B",
+ "head_kind": "tdt",
+ "expected_vocab_size": 8192,
+ "languages": V3_LANGUAGES,
+ "lang_detect": True,
+ "author": "Oruk",
+ "organization": "oruk",
+ "prefer_direct_load": True,
+ "license": "cc-by-sa-4.0",
+ "license_name": "Creative Commons Attribution-ShareAlike 4.0",
+ "license_link": "https://creativecommons.org/licenses/by-sa/4.0/",
+ },
# v2: 0.6B English-only TDT.
"parakeet-tdt-0.6b-v2": {
"variant": "tdt-0.6b-v2",
@@ -523,8 +539,6 @@ def load_nemo_model(model_spec: str, prefer_direct: bool = False):
transiently doubles disk usage; the direct path streams entries
out of the archive at near-zero transient cost.
"""
- from nemo.collections.asr.models import ASRModel
-
local = Path(model_spec).expanduser()
if prefer_direct:
@@ -535,6 +549,8 @@ def load_nemo_model(model_spec: str, prefer_direct: bool = False):
return _DirectNemoArchive(cfg, sd, sp_proto)
# else fall through to NeMo path
+ from nemo.collections.asr.models import ASRModel
+
nemo_path: Path | None = None
if local.exists():
print(f"Loading Parakeet from local path: {local}")