Parakeet-TDT-CTC-110M β€” "recordings" LinearAdapter (experimental)

A LinearAdapter (2.25M params, dim=128) on frozen nvidia/parakeet-tdt_ctc-110m, fine-tuned on a small private set of English telephone call recordings. Adapter-only (~9.5 MB); base weights are not modified or redistributed.

Results β€” honest framing

set base WER + adapter
in-domain recordings test 40.42% 38.81% (βˆ’1.61)

This adapter gives only a marginal in-domain gain and regresses off-domain β€” it is published for completeness/reproducibility, not as a strong model. For a solid telephony adapter from the same project, see the Switchboard adapter (7.0% vs 15.98% WER). Best val WER 0.4723.

Usage (NeMo)

import nemo.collections.asr as nemo_asr
m = nemo_asr.models.ASRModel.from_pretrained("nvidia/parakeet-tdt_ctc-110m")
m.load_adapters("recordings_adapter.nemo")
m.set_enabled_adapters("recordings", enabled=True)
print(m.transcribe(["call.wav"]))

Training

LinearAdapter dim=128 (in_features=512, swish, pre-norm), bf16-mixed, CosineAnnealing. Trained on private in-house telephone recordings (data not released). Base nvidia/parakeet-tdt_ctc-110m (CC-BY-4.0) β€” attribution required; this repo ships only the adapter delta.

Downloads last month
16
Inference Providers NEW
This model isn't deployed by any Inference Provider. πŸ™‹ Ask for provider support

Model tree for ozonetg/parakeet-tdt-ctc-110m-recordings-adapter

Finetuned
(8)
this model