[BUG] Windows on ARM64 (Snapdragon, Adreno X1-85): dictation returns "No speech was detected" because the Vulkan speech engine returns empty text #1312

Describe the bug

On a Windows ARM64 laptop with a Qualcomm Adreno X1 85 GPU, every dictation and file transcription ends with "No speech was detected" and 0 characters. The microphone works: the in app microphone test shows a signal, and the log shows real audio levels for each failed dictation. The speech engine receives the audio but returns empty text, with no error.

It happens with every speech model I tried (Fluid Speech Mini and Blazing Fast English), and with File Transcription of a saved recording that I confirmed contains my voice.

Dictation worked on CPU in early October on version 0.0.10 (logged as accelerator=cpu). Since mid October, the speech engine has been loading on the Vulkan backend and has produced empty text every time. I did not record exactly which update changed the default.

Reproduction steps

1. Install FluidVoice on a Windows ARM64 laptop with an Adreno X1 85 GPU.
2. Hold the dictation shortcut and speak for a few seconds, or run File Transcription on a saved recording.
3. Release the shortcut. The History entry shows "No speech was detected".

Expected behavior

Transcribed text. If the GPU path returns empty output for audio that clearly has signal, it would be good to fall back to the CPU, or to default to CPU on this hardware.

Actual behavior

Empty transcription every time. The automatic CPU fallback from the release notes never triggers, because the GPU does not report a failure; it simply returns empty text.

The AI enhancement model (fluid-1-mini) also fails on the same GPU: the log shows "ggml_backend_sched_graph_compute_async failed with error 1", then "failed to compute graph" and "llama_decode: failed to decode, ret = 2".

Workaround that fixed it

I found two environment variables in the executable and set them to cpu for my Windows user:
FLUIDVOICE_ASR_ACCELERATOR=cpu
FLUIDVOICE_LLAMA_ACCELERATION=cpu
After restarting FluidVoice, the log shows accelerator=cpu, dictation works (about 6x realtime), and AI enhancement also works on CPU (about 1.9 seconds for a 29 word dictation). This suggests the problem is specific to the Vulkan path on this GPU.

App Version

0.0.13

Windows Version

Windows 11 Home Insider Preview
version: 25H2
OS build: 26220.9836

Architecture

ARM64 (Qualcomm Snapdragon, Adreno X1-85 GPU)

Logs or crash report

Failing run (before the workaround):
[INFO] Transcribing audio: model=Fluid Speech Mini, runtime=parakeet.cpp/c-api, accelerator=auto, duration_ms=4500, samples_16k=72000
[WARN] Transcription produced empty text: duration_ms=4500, transcribe_ms=2999
[ERROR] Stop action failed: No speech was detected.

Same machine and same version after forcing CPU:
[INFO] Loaded parakeet.cpp C API backend: dll=...\parakeet\...\cpu\parakeet.dll, accelerator=cpu
[INFO] Dictation complete: model=Blazing Fast English, accelerator=cpu, words=26, audio_duration_ms=11020, transcribe_ms=2645

Native output during the failures:
ggml_vulkan: 0 = Qualcomm(R) Adreno(TM) X1 85 GPU (Qualcomm Adreno Vulkan Driver) | fp16: 1 | bf16: 0 | warp size: 64
[parakeet] pk::Backend using device: Vulkan0

Other observation: the downloaded native runtimes are the x64 builds (parakeet lib win cpu x64 and win vulkan x64) even though this is an ARM64 machine.

Screenshots or screen recording

No response