ozonetg/parakeet-tdt-ctc-110m-call-audio-adapter

Parakeet-TDT-CTC-110M — call-audio LinearAdapter

A LinearAdapter (2.25M params, dim=128) on frozen nvidia/parakeet-tdt_ctc-110m, fine-tuned on a private set of noisy conversational English telephone calls. Adapter-only (~9.5 MB); the NVIDIA base weights are not modified or redistributed. Exported adapter-only from the best training checkpoint (val WER 0.2223 @ epoch 9).

Results

set	base WER	+ adapter
in-domain call-audio (val, 28.3 h / 2,532 utt)	35.74%	22.37% (−13.37, 37% rel.)

Strong in-domain gain on messy real-world telephony. Off-domain caveat: as a domain specialist it regresses on clean/other speech (Switchboard ~18.2% vs 15.97%, LibriSpeech ~3.19% vs 2.61%). Use for noisy conversational phone audio.

Usage (NeMo)

import nemo.collections.asr as nemo_asr
m = nemo_asr.models.ASRModel.from_pretrained("nvidia/parakeet-tdt_ctc-110m")
m.load_adapters("call_audio_adapter.nemo")
m.set_enabled_adapters("call_audio", enabled=True)
print(m.transcribe(["call.wav"]))   # 16 kHz mono

Training

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

Related

• Switchboard adapter — 7.0% vs 15.98% WER on public Switchboard test.