-
Notifications
You must be signed in to change notification settings - Fork 43
Description
I just upgraded Storyteller to echogarden 2.10.1 from 2.3.1. I'm doing a test transcription with one of my books, and it fails during the tokenization step after transcription with this error:
ERROR: unreachable
err: {
"type": "Error",
"message": "unreachable",
"stack":
RuntimeError: unreachable
at wasm://wasm/00418352:wasm-function[750]:0x9cb1a
at wasm://wasm/00418352:wasm-function[1049]:0xa2616
at wasm://wasm/00418352:wasm-function[803]:0x9e16d
at wasm://wasm/00418352:wasm-function[176]:0x70ce0
at wasm://wasm/00418352:wasm-function[563]:0x95bc0
at Tiktoken.decode (/app/.next/standalone/node_modules/tiktoken/lite/tiktoken_bg.cjs:315:18)
at Whisper.tokensToText (/app/.next/standalone/node_modules/echogarden/src/recognition/WhisperSTT.ts:1746:39)
at Whisper.tokenToText (/app/.next/standalone/node_modules/echogarden/src/recognition/WhisperSTT.ts:1732:15)
at parseResultObject (/app/.next/standalone/node_modules/echogarden/src/recognition/WhisperCppSTT.ts:289:30)
at async ChildProcess.<anonymous> (/app/.next/standalone/node_modules/echogarden/src/recognition/WhisperCppSTT.ts:206:33)
}
I'm using:
- echogarden v2.10.1
- whisper.cpp v1.8.2 (with hipBLAS/ROCm, could not reproduce with CPU)
- large-v3-turbo-q5_0
Here's the file in question: https://nextcloud.friedmans.us/index.php/s/4qSntSMeRa6Ln6f
This is probably going to be a pain to reproduce, let me know if there's anything I can do to provide more information. I can reproduce it consistently with the file and configuration above.
I tried using the unquantized large-v3-turbo model and it succeeded on this file but then failed with the same error on a later file, so it seems like whatever triggers this error case is this is not totally uncommon.
I have so far not been able to reproduce (I tried all 50 chapters of this book) with a CPU build, but I am only trying with tiny on CPU.