fix: GCC 16 build, and TDT beam search failing on parakeet-tdt-0.6b-v3 - #74
Merged
Merged
Conversation
GCC 16 no longer pulls it in transitively, so the build failed. Assisted-by: Claude:claude-opus-5-5 [Claude Code]
…o score change tdt_beam_search required every zero-duration expansion to strictly lower the hypothesis score. For a near-certain label the float32 log-prob (~-6e-8) is below the resolution of a running score of a few units (spacing ~2.4e-7), so score + logp == score and the search threw "zero-duration expansion did not reduce score". parakeet-tdt-0.6b-v3 hits this on tests/fixtures/speech.wav at beam 4 and on two_speakers.wav at beam 2 and 4; the 110m anchor used in CI never does. Only an increase breaks the search, so check for that. test_transcribe_0_6b now also runs beam 4 on speech.wav and checks the top hypothesis matches the greedy reference. Assisted-by: Claude:claude-opus-5-5 [Claude Code]
mudler
approved these changes
Sep 29, 2026
mudler
approved these changes
Sep 29, 2026
EnacheB
added a commit
to EnacheB/hyprwhspr
that referenced
this pull request
Sep 29, 2026
On 98 real dictation recordings (134 chunks), beam 4 changed 34 chunks
against greedy, mostly dropping fillers and the fragments greedy decoding
invents from noise ("Um range vanquisher title.", "managing your own
investment."). It costs ~0.2 s per 10 s of audio on a Radeon 780M.
parakeet.cpp v0.5.0's beam search throws "zero-duration expansion did not
reduce score" on 28 of those 134 chunks; mudler/parakeet.cpp#74 fixes it.
Until a release carries that fix, the pinned v0.5.0 runtime fails about
one chunk in five.
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
2 participants
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.
Two small, independent fixes.
1. Build fails on GCC 16.
src/diarization_encoder.cppusesstd::copy_nwithout including<algorithm>; GCC ≤ 15 pulled it in transitively, GCC 16 does not. One-line include.2. TDT beam search throws on the v3 model. On master,
parakeet-cli transcribe --beam-size 4fails on the repo's owntests/fixtures/speech.wavwith parakeet-tdt-0.6b-v3 (and ontwo_speakers.wavat beam 2 and 4):The check required every zero-duration expansion to strictly lower the hypothesis score. A score is a sum of log-probs, so it can never rise, but for a near-certain label the log-prob (~-6e-8) is below float32 resolution at a running score of a few units (spacing ~2.4e-7), so
score + logp == score. That is correct behaviour, and the pruning only relies on scores not increasing, so the check now rejects only an increase. The 110m anchor model used in CI never produces such a label, which is why this went unnoticed.Test:
test_transcribe_0_6bnow also runs beam 4 onspeech.wavand checks the top hypothesis equals the greedy reference. It fails on master with the error above and passes with the fix. (It skips unlessPARAKEET_TEST_GGUF_06Bis set, like the existing greedy check.)Tested: full ctest with the 110m model;
test_transcribe_0_6bwithtdt-0.6b-v3-q8_0frommudler/parakeet-cpp-gguf.🤖 Generated with Claude Code