mars(kittentts): local STT add-on — faster-whisper in the audio side-venv #8

Merged
darman merged 1 commits from feat/mars-hermes-kittentts-stt into master 2026-09-20 17:24:42 +02:00
Collaborator

Stacked on #7 (base = its branch while that PR is open; retarget to master on merge).

Telegram voice-note transcription on mars, no cloud, CPU-only:

  • faster-whisper 1.2.1 (ctranslate2, CPU) installed into the SAME kittentts side-venv via a new `stt-requirements.txt`, appended by the existing provisioning unit; input-hash stamp extended.
  • Second auto-patchelf pass RPATHs the newly installed ctranslate2/av binaries.
  • Container env gains `PYTHONPATH=/opt/data/kittentts-venv/lib/python3.13/site-packages` so hermes's local-STT provider imports faster-whisper in-process — the image venv is read-only for uid 986, so the built-in lazy installer can never work there.
  • Validated live: a real Telegram voice note (4.2 s Ogg/Opus) transcribed on CPU: `lang: en (p=0.52), "Hi Luna, great to hear you, can you hear me too?"` — exact match, ~seconds on i5-4460.

Model note: the whisper base model (~150 MB) downloads on first use via the hub; with HF_HOME already pointing at the presaged.bytes cache dir it persists across boots. If you want the cache pre-seeded/ro like the TTS model, say the word and I'll extend the staging unit.

Stacked on #7 (base = its branch while that PR is open; retarget to master on merge). Telegram voice-note transcription on mars, no cloud, CPU-only: - faster-whisper 1.2.1 (ctranslate2, CPU) installed into the SAME kittentts side-venv via a new \`stt-requirements.txt\`, appended by the existing provisioning unit; input-hash stamp extended. - Second auto-patchelf pass RPATHs the newly installed ctranslate2/av binaries. - Container env gains \`PYTHONPATH=/opt/data/kittentts-venv/lib/python3.13/site-packages\` so hermes's local-STT provider imports faster-whisper in-process — the image venv is read-only for uid 986, so the built-in lazy installer can never work there. - Validated live: a real Telegram voice note (4.2 s Ogg/Opus) transcribed on CPU: \`lang: en (p=0.52), "Hi Luna, great to hear you, can you hear me too?"\` — exact match, ~seconds on i5-4460. Model note: the whisper base model (~150 MB) downloads on first use via the hub; with HF_HOME already pointing at the presaged.bytes cache dir it persists across boots. If you want the cache pre-seeded/ro like the TTS model, say the word and I'll extend the staging unit.
darman changed target branch from feat/mars-hermes-kittentts to master 2026-09-20 17:24:31 +02:00
darman added 1 commit 2026-09-20 17:24:31 +02:00
Telegram voice-note transcription on mars, no cloud: faster-whisper 1.2.1
(ctranslate2 CPU) installed into the SAME kittentts side-venv via
stt-requirements.txt (appended by the provisioning unit; input stamp
extended). Second auto-patchelf pass RPATHs the newly installed
ctranslate2/av binaries. Container env gains
PYTHONPATH=/opt/data/kittentts-venv/... so hermes's local-stt provider
imports faster-whisper in-process (the image venv is read-only for uid
986, so the built-in lazy installer cannot work there).

Validated live: Telegram voice note transcribed on CPU (i5-4460,
lang=en, text exact) with the same pins.
darman merged commit 71900ddfcf into master 2026-09-20 17:24:42 +02:00
darman deleted branch feat/mars-hermes-kittentts-stt 2026-09-20 17:24:42 +02:00
Sign in to join this conversation.
No Reviewers
No labels
2 Participants
Notifications
Due Date
No due date set.
Dependencies

No dependencies set.

Reference: darman/homelab#8