Spaces:
Running
Running
Commit History
16 days ago
-
Brave: stop recordings turning into blanks (getChannelData farbling) Running
Verified signatureBrave's fingerprinting protection rescales an AudioBuffer in place on every getChannelData() call (x0.99982 per call on Brave 1.94). Gradio's audioBufferToWav calls it once per sample, so the uploaded recording decayed to digital zeros after ~1 s: the speech waveform was replaced by blanks, the model got silence, and encoding 5 s took ~15 s. The page now caches getChannelData per buffer (same live array in every browser; installed only when a probe shows reads mutate data). Measured, same fake-mic input: Brave upload non-zero 4% -> 79% (Chromium: 79%), Stop->transcript 1.1 s. Server logs and notes trailing digital silence so a recurrence is visible.cd19be5 -
Mic: stop recording blanks at start, close leaked mic streams, honour device dropdown
Verified signatureGradio 5.44's recorder opens the mic on Record and starts MediaRecorder immediately, so the mic start-up (1.8 s measured on Chromium; 1-2 s typical for Bluetooth headsets) was recorded as exact digital zeros and the first words were lost. A getUserMedia wrapper now waits for real samples before recording starts (timer appears when the mic is live). Gradio also opens the mic three times per recording and stops one, leaving two streams live after Stop; the wrapper hands out one stream per recording. The input-device dropdown, ignored by Gradio, is honoured. Server side, an all-zero recording gets an explicit message and >=0.5 s of leading digital silence gets a note. Measured (Chromium fake mic): leading silence 1.83 s -> 0.00 s; live mic tracks after Stop 2 per recording -> 0.987b192 -
Labelled examples with references, 60 s truncation, gated-model error, corrected WER
Verified signature- examples: 4 NepTel call segments (canonical sha256) + 2 FLEURS read-speech clips, each with its human-reviewed reference; call clips also show the published Whisper-large-v3 output - recordings/uploads over 60 s are cut to the first 60 s with a note instead of refused - startup raises a clear error when the gated model cannot be downloaded (HF_TOKEN secret) - description/README: 36.3% WER (reproducible decode) and Whisper 96.3%, replacing the retracted 33.8% and ~99%; measured CPU time ~5-7 s per 25 s clipac18673
-
examples: run_on_click so clicking a sample transcribes it
Verified signatureee038c5 -
auto-transcribe on stop_recording: submitting mid-recording silently sent nothing
Verified signatureb8d8ebe -
fix mic input: always downmix to mono 16 kHz (stereo recordings crashed NeMo)
Verified signature145c9e7 -
restore handler: ffmpeg fallback for mic webm, show_error
Verified signature4d51cdd -
temp: diagnostic handler
Verified signature0bb4d49 -
revert format=wav (broke preprocess); convert mic webm via ffmpeg in handler; show_error
Verified signaturec12a0a7 -
fix mic recording: disable SSR (0.0.0.0 root-url bug), force wav format
Verified signatureed8fed2
-
Upload folder using huggingface_hub
Verified signatureb86a6d6 -
Upload folder using huggingface_hub
Verified signature48db520 -
Upload folder using huggingface_hub
Verified signature783c42e -
initial commit
Verified signature546a765