Error when input file is longer than 5min

Transcription with parakeet is working ok with small files, but with files a longer than 5min it gives the following error: ErrorOutOfDeviceMemory.

I'm using the vulkan version.

Steps to reproduce

I have a 2hr+ recording that want to transcribe, gives error and did try with 10min, like this:

$ ffmpeg -t 00:10:00 -i in.wav -c copy p1.wav

$ ffprobe p1.wav
Input #0, wav, from 'p1.wav':
  Metadata:
    encoder         : Lavf62.12.102
  Duration: 00:10:00.00, bitrate: 705 kb/s
  Stream #0:0: Audio: pcm_s16le ([1][0][0][0] / 0x0001), 44100 Hz, 1 channels, s16, 705 kb/s

$ parakeet-cli transcribe --model $HOME/00_M/tdt-0.6b-v3-q4_k.gguf --input p1.wav --json > out.json
ggml_vulkan: Found 1 Vulkan devices:
ggml_vulkan: 0 = NVIDIA GeForce GTX 1660 SUPER (NVIDIA) | uma: 0 | fp16: 1 | bf16: 0 | warp size: 32 | shared memory: 49152 | int dot: 1 | matrix cores: none
[parakeet] pk::Backend using device: Vulkan0
ggml_vulkan: Device memory allocation of size 900240064 failed.
ggml_vulkan: vk::Device::allocateMemory: ErrorOutOfDeviceMemory
ggml_gallocr_reserve_n_impl: failed to allocate Vulkan0 buffer of size 16698365856
[parakeet] Backend::compute: ggml_gallocr_alloc_graph failed
Segmentation fault         (core dumped) parakeet-cli transcribe --model $HOME/00_M/tdt-0.6b-v3-q4_k.gguf --input p1.wav --json > out.json

• I also did try the pipe example with ffmpeg, gives same error.
• If I do the same, but with 5min, works.
• I can do the the full 2hr+ audio with Whisper without any problem, but I'm trying parakeet now, because is faster.
• I can chop the audio in many 5 min files, but is more like a workaround than a solution, also gives problems to join subtitles.
• I did use frequently llama.cpp and sometimes sd.cpp, if it counts as experience.
• My system is updated.

System

distro: Arch Linux x86_64
kernel: 6.18.25-1-lts
cpu: Intel i7-4790 (8) @ 3.600GHz
ram: 32 GB
gpu: NVIDIA GeForce GTX 1660 SUPER (6GB VRAM)

Start with the parakeet-cli transcribe entry point and reproduce the difference between the working 5-minute input and the failing 10-minute input using the commands in the issue. Inspect the Vulkan allocation and compute path around the reported ggml_gallocr and Backend::compute failures. Done means longer audio no longer exhausts device memory or segfaults, while preserving transcription output.