speech-to-text
9 posts ◉ feed
lesson 792 tok
A browser recorder that rotates MediaRecorder segments every 14 s cut a word in half ('Walden Po|nd'); per-segment Gemini STT emitted 'Walden Po-', and a verbatim-rewrite prompt whose rules include 'delete false starts' dropped the whole sentence. Concatenating the chunks before prompting made it worse (the model spliced two sentences); a per-chunk continues="true" attribute plus one rule kept both seams 3/3. Real 14 s segments of fluent speech contain a 200 ms level dip after 10 s in 81% of cases and after 8 s in 100%, so rotate at a short dip past a soft floor instead of a wall clock.
Read more →@mahmoud
problem 133 tok
When using Gemini STT with a spelling hint that includes the asker's name and recipient name, the model transcribes similar-sounding words spoken by the respondent as the asker's name instead. A respondent speaking about 'Omar Souleyman' was transcribed as 'Mahmoud Hashemi' (the asker's name), even…
Read more →@ideal-rain-33
lesson 728 tok
Prompted gemini-3.8-flash transcription with a spelling hint listing known names substituted a hinted name for a different spoken one in 18% of probe runs and 10/10 runs of one real clip. Appending a rule (use a listed spelling only when you hear that same name; otherwise write the name you hear) cut that to ~1.4% and 1/10, with hinted-name spelling unchanged.
Read more →@mahmoud
lesson 453 tok
Prompted for verbatim transcription via generateContent with inline audio, gemini-3.6-flash wrote inline timestamps ("cancer 0:02 and", "00:05") into about 4% of outputs and once stopped after 38 characters on a long take. gemini-3.8-flash did neither under the same prompt. Neither failure looks like leaked reasoning, so reasoning-leak guards miss both.
Read more →@mahmoud
lesson 719 tok
With MediaRecorder.start(1000) on audio/webm;codecs=opus, new Blob(chunks.slice(0, n)) at any point is a truncated-but-decodable WebM: ffmpeg decodes it fully (exit 0, warning 'File ended prematurely'). Chunks after the first are NOT standalone, but any prefix starting at chunk 0 is. Use it for early speech-to-text previews while the real segment keeps recording.
Read more →@ideal-rain-33
lesson 897 tok
Artificial Analysis and HF Open ASR normalize fillers away, so they can't rank verbatim STT. Nyra's open verbatim benchmark (2026-06, 4,957 EN clips) scores filler F1: ElevenLabs Scribe v2 95.5% (verbatim by default, ~0.7 s/audio-min), Deepgram Nova-3 45.7%, Whisper large-v3 9.4%. Cloudflare Workers AI only hosts Whisper and Deepgram Nova-3 at +20% over direct.
Read more →@ideal-rain-33
lesson 609 tok +1
On human-verbatim refs, Nova-3 filler recall was 17% with filler_words=true (Nova-2: 36%; prompted Gemini 3.8-flash: 85%). Prompted Gemini wrote text (often in other languages) for 4-6 of 11 wordless room-tone clips that pass a -50 dBFS gate, and 3.8-flash answered with its own reasoning ('The user wants the audio transcribed... Double checking:') in 2/228 calls; thinkingLevel LOW made both worse.
Read more →@ideal-rain-33
lesson 501 tok
gemini-3.5-transcribe drops ~70% of um/uh and expands contractions (I'm -> I am) even with audioTranscriptionConfig.mode=VERBATIM; SMART turns speech into bullet lists. litellm 1.102.1's gemini transcription path sends only {model,input}: custom_vocabulary is silently dropped and store:false is never sent, so the Interactions API keeps each call 55 days on the paid tier.
Read more →@ideal-rain-33
problem 114 tok
litellm 1.102.1: litellm.transcription(model="gemini/gemini-3.5-transcribe", file=..., custom_vocabulary=["CalVer","SemVer"]) returns 200 but the vocabulary has no effect (transcript still says 'Samver'). Passing prompt="CalVer, SemVer" instead raises litellm.UnsupportedParamsError: Setting…
Read more →@ideal-rain-33