Downloads · 30 days
86
75% of all-time downloads
milanakdj/parakeet-tdt-0.6b-v3-nepali
parakeet-tdt-0.6b-v3-nepali is a automatic speech recognition model from milanakdj. Use it when you need speech turned into text. It is set up for nemo. The card lists the license as cc-by-4.0.
This model does not work. Test WER is 0.9519 — about 19 words wrong in 20. It is published as a reproducible negative result and a baseline to measure future runs against. Do not use it for transcription.
Downloads · 30 days
86
75% of all-time downloads
All-time downloads
114
Public
Repo size
2.5 GB
Likes
0
Public
Click a slice to open those files.
.nemo2.5 GB · 100%
From the Hugging Face model README
This model does not work. Test WER is 0.9519 — about 19 words wrong in 20. It is published as a reproducible negative result and a baseline to measure future runs against. Do not use it for transcription.
For working Nepali ASR on this corpus, use whisper-large-v3, which reaches 0.1145 WER on the same audio.
nvidia/parakeet-tdt-0.6b-v3 fine-tuned on Nepali by extending its
SentencePiece vocabulary. v3's 8192 pieces cover 25 European languages and
contain zero Devanagari, so 436 Nepali pieces were appended after them,
keeping every original id stable, and the pretrained embedding and joint rows
were copied back by slice. The encoder was frozen for 2000 steps, then trained
at a reduced rate.
| WER | |
|---|---|
| test (held-out) | 0.9519 |
| best validation | 0.9360 (step 16,408) |
| whisper-large-v3, same corpus | 0.1145 |
Validation converged at step 16,408 and did not improve across the remaining 23,000 steps.
| Data | lilgoose7777/slr-combined-nepali-tts2, 166,967 rows |
| Speech | 90.2 hours (114.8 h of files after silence trimming) |
| Vocabulary | 8192 → 8628, fertility 3.18 tokens/word, 0 unknown |
| Steps | 40,000 optimizer steps, 7 epochs |
| Batch | 2 × 16 accumulation = effective 32 |
| LR | 4e-4 decoder/joint, 1e-4 encoder, cosine, 1000 warmup |
| Precision | bf16-mixed |
| Hardware | NVIDIA DGX Spark (GB10), 18 GB cap on a shared 121 GB pool |
offset fields moved WER off a hard 1.05 plateau for the first
time, and flipped the error mode from insertions to deletions.Single script, four steps: smoke, manifests, tokenizer, train.
import nemo.collections.asr as nemo_asr
m = nemo_asr.models.ASRModel.restore_from("parakeet-tdt-0.6b-v3-nepali.nemo")
print(m.transcribe(["clip.wav"]))