Error maximum size and transcribing #104

Activity

1. christoic commented on Jul 22, 2024

   

   christoic

   on Jul 22, 2024

   Author

   More actions

   Erratum : work with my android phone...not with my Windows 10 PC
2. chhoumann commented on Jun 15, 2026

   

   chhoumann

   on Jun 15, 2026

   Owner

   More actions

   Thanks for the detailed report — and sorry for the long delay. I went back through this against current master and the "maximum size" failure you hit has since been fixed, by two independent changes that landed after this issue was filed. Closing as resolved, with evidence below.

   What was going wrong (Jul 2024)

   Your report had two symptoms, with two different root causes:

   1. The storage file has been changed but exceeds the maximum size. Skipping: [path] — this is an Obsidian Sync message. The old transcription flow downloaded the full episode into the vault (vault.createBinary) before transcribing, so Sync tried to sync a large audio file and rejected it.
   2. [Error transcribing chunk 0] [Error transcribing chunk 1] [Error transcribing chunk 2] — the fallback strings emitted when every OpenAI request failed. The whole, undivided episode was sent in one request, and OpenAI's Whisper API rejects any single request whose audio file exceeds 25 MB.

   Why it's fixed on current master

   1. Every OpenAI request is now capped at 20 MB (under the 25 MB limit).
   Audio is split into chunks before transcription. The cap is CHUNK_SIZE_BYTES = 20 * 1024 * 1024 (src/services/TranscriptionService.ts:56), and it's enforced on the only request site, client.audio.transcriptions.create (src/services/TranscriptionService.ts:434), which is fed exclusively by createChunkFiles:

   • Binary path — createBinaryChunkFiles (:219-248): returns a single file only when the buffer is <= 20 MB, otherwise slices at exact 20 MB boundaries.
   • m4a → WAV path — renderWavChunks (:288-326): bounds each WAV file at WAV_HEADER_SIZE + sampleCount * blockAlign <= 20 MB, including the PCM expansion. (Mono/stereo/4-channel all compute to ≤ 20,971,520 bytes.)
   • If WAV decoding fails, it falls back to the 20 MB binary chunker (:204-211), so no path bypasses the cap.

   So no single request can exceed ~20 MB — about 4 MB under the 25 MB limit, which comfortably covers multipart overhead. Chunking was introduced in #128.

   2. Transcription no longer writes the audio into the vault.
   The transcription flow now fetches audio into memory via getEpisodeAudioBuffer (src/downloadEpisode.ts:397): "the episode's own stream URL is fetched into memory. Nothing is written to the vault" (:394-395). The only vault.createBinary call (src/downloadEpisode.ts:223) lives in createEpisodeFile, used solely by the Download command — not by transcription. So the Sync size-rejection from symptom (1) can no longer be triggered by transcribing. This landed in #182 (the fix for #107).

   Verification

   • TranscriptionService unit tests pass (11/11), including the chunk-splitting tests.
   • I ran an adversarial review specifically trying to construct an input (large/stereo/high-sample-rate episode, decode-failure fallback, un-chunked path) that still produces a >25 MB request or a transcription-time vault write — none exists.

   On your OpenRouter / API-path request

   Btw is it possible to add the API path as well in the settings as I am using https://openrouter.ai/

   That's a great idea, but it's a separate enhancement from this "max size" bug — the OpenAI client is currently constructed with no configurable baseURL (src/services/TranscriptionService.ts:530), so an OpenRouter key won't work today regardless of file size. There isn't an open issue for it yet; if you'd still like a configurable transcription endpoint, please open a dedicated feature request and we'll track it there. (Note: OpenRouter would also need to expose a Whisper-compatible audio/transcriptions endpoint for it to work end-to-end.)

   Thanks again for the kind words and the report. 🙏
3. chhoumann

   closed this as completedon Jun 15, 2026