Per-model Spark speech recipes on a source-built audio backend - #41
Merged
Merged
Conversation
Replace the host-local prebuilt media image with a locally built lloom/spark-audio:source-<sha256> image and one public recipe per model: Qwen3-TTS 1.7B CustomVoice, Base (voice cloning) and VoiceDesign, and Whisper large-v3-turbo. Each recipe runs its checkpoint in its own managed container with a loopback port, a read-only mount of only that model's directory, and offline Hugging Face mode. Downloads are pinned to immutable revisions with per-file SHA-256. openai-whisper cannot load the Transformers-format repository, so the Whisper recipe fetches OpenAI's original large-v3-turbo.pt from a byte-identical Hugging Face copy whose hash matches openai-whisper v20250625. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Replace the third-party Hugging Face copy with an unmodified mirror of OpenAI's large-v3-turbo.pt; the enforced SHA-256 is unchanged. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
# Conflicts: # CHANGELOG.md # docs/recipes.md # package.json
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
Replaces a host-local recipe that ran four speech models from one 32.8 GB prebuilt image with a source-built backend and one recipe per model. Stacked on #37.
backends/spark-audio/install.pybuilds or reuseslloom/spark-audio:source-<digest>, the same pattern as the other source-built backends. The digest covers only the Dockerfile and the server. The backend is registered inbackends/catalog.json.linux-nvidia-spark-audio-qwen3-tts-1-7b-customvoicelinux-nvidia-spark-audio-qwen3-tts-1-7b-base(voice clone)linux-nvidia-spark-audio-qwen3-tts-1-7b-voicedesignlinux-nvidia-spark-audio-whisper-large-v3-turboWhisper weights source
The adapter loads Whisper with openai-whisper, which needs OpenAI's original
large-v3-turbo.pt.openai/whisper-large-v3-turboonly publishes Transformers weights. The recipe downloads the file fromdataangel/whisper-large-v3-turbo-openaiat a pinned revision. That repo is an unmodified mirror copied from OpenAI's download URL, with Whisper's MIT license. The sha256 is enforced and matches the one openai-whisper v20250625 publishes.Test plan
node test/spark-audio-recipes.test.mjs: each recipe alone, and all four together in both orders, get independent runtimes and ports, the pinned image, one read-only mount, and no host pathsnode --test test/speech-stream.test.mjs(12 pass),node test/tts-catalog.test.mjspytest backends/spark-audio(53 pass)npm run check,npm run test:unit,node scripts/check-package.mjs🤖 Generated with Claude Code