Load with 256-frame local attention by default

README.md

@@ -15,7 +15,7 @@ tags:
 
 # parakeet-ft-v1
 
-A fine-tune of [NVIDIA Parakeet TDT 0.6B v3](https://huggingface.co/nvidia/parakeet-tdt-0.6b-v3) for long recordings. v3 sometimes skips whole passages of a long recording, and this model almost never does: on 16 recordings kept out of training, v3 skipped 87 passages and this model 3. Same architecture, same tokenizer, same 0.6B parameters, loadable with [NeMo](https://github.com/NVIDIA-NeMo/Speech) exactly like the original.
+A fine-tune of [NVIDIA Parakeet TDT 0.6B v3](https://huggingface.co/nvidia/parakeet-tdt-0.6b-v3) for long recordings. v3 sometimes skips whole passages of a long recording, and this model almost never does: on 16 recordings kept out of training, v3 skipped 87 passages and this model 3. Same architecture, same tokenizer, same 0.6B parameters, loadable with [NeMo](https://github.com/NVIDIA-NeMo/Speech) like the original. It loads with the local attention it was trained with, so long recordings fit in memory without extra settings.
 
 It was trained only on its own output. No person wrote or corrected a transcript, and no other speech recognizer was involved.
 
@@ -50,7 +50,6 @@ With plain NeMo:
 import nemo.collections.asr as nemo_asr
 
 asr = nemo_asr.models.ASRModel.from_pretrained("freetranscribe/parakeet-ft-v1")
-asr.change_attention_model("rel_pos_local_attn", [256, 256])
 print(asr.transcribe(["recording.wav"])[0].text)
 ```
 
@@ -103,7 +102,7 @@ The next table shows what each step adds on Earnings-22 (125 earnings calls, abo
 | Plus cuts at pauses, 90 seconds at most | 11.32 | 10.36 |
 | Plus 5 seconds of overlap | 11.34 | 10.27 |
 
-Wider attention and cutting help both models, and cutting helps v3 far more. The overlap helps only the fine-tuned model: on v3 the last two rows are within run-to-run variation. The wider attention is a one-line change you can make in plain NeMo.
+Wider attention and cutting help both models, and cutting helps v3 far more. The overlap helps only the fine-tuned model: on v3 the last two rows are within run-to-run variation. This model loads with 256-frame attention already. On v3 it is one line in plain NeMo: `asr.change_attention_model("rel_pos_local_attn", [256, 256])`.
 
 I chose the settings on Earnings-22 with the fine-tuned model and 256-frame attention, without overlap unless stated:
 

parakeet-ft-v1.nemo

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:9451511f71a79fd3bb536980d2654911f4033b0174c456320abc4c4845f4e815
+oid sha256:c717d4dac2ffa9901278974a4a80ba03ba10045712197897201439a78d0c7c86
 size 2509322240