ozonetg/parakeet-tdt-ctc-110m-switchboard-adapter

Parakeet-TDT-CTC-110M — Switchboard LinearAdapter

A lightweight LinearAdapter (2.25M trainable params, dim=128) fine-tuned on top of the frozen nvidia/parakeet-tdt_ctc-110m for conversational telephone speech. The base model's 114.6M weights are untouched — this file is only the ~9.5 MB adapter.

Results — Switchboard test

model	WER	utterances	audio
base parakeet-tdt_ctc-110m	15.98%	50,165	63.4 h
+ this adapter	7.00%	50,165	63.4 h

−8.98 WER absolute / 56% relative on the public hhoangphuoc/switchboard test split (650,491 reference words; 23,703 sub / 9,284 ins / 12,577 del).

Off-domain caveat (important): this is a telephony specialist. On clean read speech it regresses (LibriSpeech WER ~4.96% vs ~2.61% base). Use it for conversational/telephone audio, not general ASR.

Usage (NeMo)

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

Training

• Base: nvidia/parakeet-tdt_ctc-110m (frozen, 114.6M params)
• Adapter: LinearAdapter, dim 128, 2,245,632 trainable params
• 50 epochs, effective batch 32, LR 5e-4, CosineAnnealing (500 warmup), bf16-mixed, grad-clip 1.0; best val WER 0.0687 @ epoch 4; ~6.8 h on one GPU.
• Data: Switchboard (LDC; not redistributed here).

License / attribution

Adapter weights: CC-BY-4.0. Requires the base model nvidia/parakeet-tdt_ctc-110m (NVIDIA, CC-BY-4.0) — load it from NVIDIA and apply this adapter on top. This repo does not redistribute NVIDIA's weights.