FocsleFocsle
OE

Whisper-Edge-Tiny

8%
by OpenAI-Edge

Edge fork of Whisper-Tiny with NPU-friendly attention rewriting. INT4 variant runs on the Pi alone.

Embedded ASRMITINT8INT4asrenglishtiny
412K downloads 25K deploymentsUpdated Mar 18, 2028
Headline:220ms · Raspberry Pi 5 + Hailo HAT · INT8

Cross-chip benchmark matrix

Every supported chip, in matched-pair runs from the Fo’c’sle HIL lab. Sortable by any column — click a header. Cells where the chip can’t run this model show Not supported.

Chip platformQuantLatency p50(ms)Latency p95(ms)ThroughputAcc. retention(%)Power(W)Memory(MB)Tested by
Apple Neural Engine (M4)
38 TOPS · SoC
INT8203.5242.831 tok/s97.76.846Fo’c’sle HIL
Q
Snapdragon 8 Gen 3 NPU
45 TOPS · SoC
INT8216.0260.630 tok/s97.66.846Fo’c’sle HIL
Q
Qualcomm QCS8550
48 TOPS · SoC
INT8221.7262.429 tok/s97.313.346Publisher
H
Hailo-8
26 TOPS · M.2
INT8355.9446.118 tok/s98.12.146Publisher
π
Raspberry Pi 5 + Hailo HAT
26 TOPS · HAT
INT8368.8461.217 tok/s97.76.946Fo’c’sle HIL
Q
Qualcomm QCS6490
12 TOPS · SoC
INT8526.3680.212 tok/s98.28.046Fo’c’sle HIL
R
Rockchip RK3588
6 TOPS · SoC
INT8836.51063.07.7 tok/s97.59.746Fo’c’sle HIL
G
Google Coral Edge TPU
4 TOPS · USB
INT81008.91277.66.3 tok/s98.41.746Fo’c’sle HIL
H
Hailo-10H
40 TOPS · M.2
Not supported
N
NVIDIA Jetson Orin Nano
40 TOPS · SoM
Not supported
N
NVIDIA Jetson AGX Orin
275 TOPS · Module
Not supported
N
NVIDIA Jetson Thor
2070 TOPS · Module
Not supported
A
Ambarella CV5
16 TOPS · SoC
Not supported
A
Ambarella CV72
32 TOPS · SoC
Not supported
M
MediaTek Genio 700
4 TOPS · SoC
Not supported
I
Intel Movidius Myriad X
4 TOPS · SoC
Not supported
A
AMD Versal AI Edge VE2302
22 TOPS · SoC
Not supported
Leader per columnLibriSpeech test-clean · 16 kHz streaming

HIL conditions

All numbers measured on Fo’c’sle HIL rigs in Tel Aviv (primary), Munich (secondary), and Pittsburgh (robotics). Single-stream, batch-1, real preprocessing, real downstream consumer. p50/p95 are over 10,000-frame steady-state windows after a 30-second warm-up. Power draw is package power, not wall power. Memory footprint is the resident model + activations footprint at peak — not on-disk.

Submitted publisher numbers are accepted only if they reproduce within ±8% of an HIL-lab matched run on the same chip in the same input mode. Otherwise they live separately under the Discussion tab.