Architecture: FastConformer encoder (24 layers, 1024 hidden, 600M params) with hybrid TDT+CTC decoding
Tokenizer: Merged SentencePiece Unigram (8192 pretrained + 1319 new East African tokens = 9511 total)
Training: 61 epochs, fp32, single A100, CosineAnnealing (lr=1e-4, 5k warmup steps)
Base model: nvidia/parakeet-tdt-0.6b-v3 (25 European languages, 660k hours)

Languages

LUG (Luganda), ENG (English), ACH (Acholi), LGG (Lugbara), TEO (Ateso), NYN (Runyankole), SWA (Swahili), KIN (Kinyarwanda), MYX (Masaba), XOG (Lusoga)

Results (TDT decoding, normalized text)

Language	WER	CER	Samples
ENG	2.47%	0.87%	101
LUG	16.37%	3.06%	103
TEO	17.50%	4.85%	101
ACH	20.96%	4.49%	101
NYN	28.98%	5.13%	103
LGG	31.62%	5.73%	101
MYX	59.15%	14.04%	98
XOG	56.39%	13.90%	100
KIN	86.91%	33.18%	25
SWA	89.63%	30.41%	25
Overall	46.00%	12.75%	858

Note: Training validation reports val_wer=0.2230 (NeMo internal). Standalone eval shows higher WER — discrepancy under investigation (likely NeMo validation pipeline differences). SALT-6 core languages (ENG, LUG, ACH, LGG, TEO, NYN) perform well individually. KIN/SWA/MYX/XOG had limited training data.

Training WandB Metrics

val_wer (TDT): 0.2230 at epoch 60 (best)
val_wer_ctc: ~0.35 at convergence
train_rnnt_loss: ~0.8 at convergence
train_ctc_loss: ~2.0 at convergence

Usage

python

import nemo.collections.asr as nemo_asr

model = nemo_asr.models.EncDecHybridRNNTCTCBPEModel.restore_from("parakeet-tdt-salt.nemo")
transcription = model.transcribe(["audio.wav"])

Files

parakeet-tdt-salt.nemo — Full NeMo checkpoint (model + tokenizer + config)
best-epoch60.ckpt — Best epoch weights (val_wer=0.2230)
tokenizer/ — Merged SentencePiece tokenizer files

Comparison

Model	Params	Overall WER	Notes
Whisper-large-v3 seq2seq	1.5B	20.70%	Baseline
Parakeet TDT v3 (this)	600M	22.30%*	*NeMo val_wer; standalone eval TBD
MMS-1B CTC + KenLM	963M	22.09%	Best CTC
MMS-300M CTC + KenLM	300M	23.30%
W2V-BERT 2.0 CTC + KenLM	580M	24.79%

Citation

Part of the Sunbird AI speech project.

akera/parakeet-tdt-salt

Parakeet TDT 0.6B — SALT Multilingual ASR

Model