From 27331d66e4ab83695cd8b3481a14223efd5f152c Mon Sep 17 00:00:00 2001 From: archipelago Date: Wed, 22 Jul 2026 15:51:32 -0400 Subject: [PATCH] =?UTF-8?q?perf(pine):=20whisper=20--beam-size=201=20?= =?UTF-8?q?=E2=80=94=20~45%=20faster=20STT,=20same=20transcripts?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit Image default is beam 5 on x86 (1 on ARM). Benchmarked on framework-pt (i5-1135G7, base-int8, real speech): beam 1 transcribes identical text ~45% faster; extra CPU threads made it slower, so only the beam changes. App revision 3.4.2 > image 3.4.1 so catalog-driven nodes roll the args change; a 3.4.1-1 suffix would semver-compare LOWER and never ship. Co-Authored-By: Claude Fable 5 --- apps/pine-whisper/manifest.yml | 12 ++++++++++-- 1 file changed, 10 insertions(+), 2 deletions(-) diff --git a/apps/pine-whisper/manifest.yml b/apps/pine-whisper/manifest.yml index 57e39490..9951e7d7 100644 --- a/apps/pine-whisper/manifest.yml +++ b/apps/pine-whisper/manifest.yml @@ -1,7 +1,11 @@ app: id: pine-whisper name: Pine Whisper (STT) - version: "3.4.1" + # App revision 3.4.2 = upstream wyoming-whisper 3.4.1 image + tuned args + # (--beam-size 1). Bumped past the image version so catalog-driven nodes + # pick up the args change; the pre-release form "3.4.1-1" would compare + # LOWER than 3.4.1 under semver and never roll out. + version: "3.4.2" description: Wyoming-protocol faster-whisper speech-to-text engine. Internal Pine voice-assistant stack member — turns speech captured by a PineVoice satellite into text for Home Assistant Assist. category: home @@ -18,7 +22,11 @@ app: network_aliases: [pine-whisper] # The image entrypoint already binds tcp://0.0.0.0:10300; these args only # pick the model + language (mirrors the pine ha-stack.yml compose command). - custom_args: ["--model", "base-int8", "--language", "en"] + # --beam-size 1: the image default is 5 on x86 (1 on ARM). Benchmarked on + # framework-pt (i5-1135G7, base-int8): beam 1 transcribes the same text + # ~45% faster — the standard low-latency setting for short voice commands + # (HA's own whisper add-on defaults to 1). + custom_args: ["--model", "base-int8", "--language", "en", "--beam-size", "1"] dependencies: - storage: 2Gi