Spaces:
Paused
Paused
Commit History
-
fix(auth): strip whitespace from SARVAM_API_KEY secret
3b1f1bb -
fix(sarvam): streamline Sarvam STT to REST API and format translate response as JSON
04e58e4
-
fix: route Azure, Sarvam and NLLB codes through the registry
Each engine had its own small hardcoded code map with a silent fallback, so an unsupported language produced confidently wrong output instead of an error: - Sarvam STT mapped unknown languages to "en-IN", transcribing e.g. German audio as Indian English. - Sarvam translate defaulted the source to "en-IN" and the target to "hi-IN", so an unsupported pair silently returned Hindi. - Azure passed unmapped codes through raw, handing it "de" where it expects "de-DE". - NLLB mapped 11 short codes and passed the rest through, so "ar" reached the model instead of "ara_Arab". All four now resolve through languages.py, covering 46 languages, and the Sarvam paths reject unsupported languages with a 400 rather than guessing. The frontend already restricts Sarvam to Indic languages, but the API is public and that gate can regress. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>97806b5 -
feat: derive PDF fonts from a 46-language registry
The PDF exporter used a hardcoded 8-language font map keyed mostly by FLORES-200 codes, falling back to Latin-only NotoSans for anything else. Two consequences: - 15 of the 46 offered languages had no usable font and rendered as empty boxes (Arabic, Hebrew, Thai, Bengali, Gujarati, Gurmukhi, Korean and the Indic set). - The map had "tel_Telu" but not the bare "te". The frontend sends the short code as src_lang and the FLORES code as tgt_lang, so exported PDFs rendered the translation correctly while the source transcript came out blank for every Indic language. Both now resolve through languages.py, which maps language -> script -> font, so adding a language is a single row. The Dockerfile fetches one face per script (15) and the build fails if any is missing, rather than degrading silently at runtime. Text shaping was keyed off `!= "en"`, enabling it for CJK and Latin where it does nothing; it now follows the script. LANGUAGE_PROMPTS is deliberately not taken from the re…157d9bc
-
fix: use force flag for symlinks to avoid build collision
c7faad7 -
fix: update Dockerfile to CUDA 12 base and fix GPU device typo
4643c0a
-
feat: smart task cancellation on client disconnect
HARI PREETHAM85f8e7f -
fix: pure telugu language lock and beam_size=5 restoration
HARI PREETHAM5cc4545 -
feat: crystal-clear audio pipeline with gapless recording and normalization
HARI PREETHAM6022cb9 -
feat: high-accuracy telugu dictionary and tuned syllable padding
HARI PREETHAMc8575ce -
fix: aggressive hallucination suppression and decoding hardening
HARI PREETHAMc92166e -
fix: implement translation hallucination guard and harden whisper vad
HARI PREETHAMff1cf8e -
feat: consolidate translation engine to NLLB-1.3B and remove IndicTrans2
HARI PREETHAM1cf40b8 -
feat: unlock nvidia t4 gpu for large-v3 transcription
HARI PREETHAM501f175
-
fix: explicit 8-vCPU optimization for non-GPU hardware
HARI PREETHAMaa3ba56 -
fix: optimize beam size for speed and remove hallucination-prone prompt
HARI PREETHAM41f6cec -
fix: relax confidence threshold and refine hallucination regex
HARI PREETHAM2268154 -
feat: switch back to whisper-large-v3-turbo
HARI PREETHAM4458b5e -
feat: switch to whisper-medium for increased speed
HARI PREETHAMa335010 -
feat: high-impact whisper decoding upgrade (turbo + beam 8 + filters)
HARI PREETHAM7bd0494 -
debug: add VAD logging for calibration
HARI PREETHAM6948547 -
fix: eliminate repetitive hallucinations and harden VAD
HARI PREETHAMd0ce3a2 -
feat: quality-focused large-v3 for Telugu and hallucination guard
HARI PREETHAM4afa6ff -
feat: stable optimized backend with large-v3-turbo and loop prevention
HARI PREETHAMcd14d03 -
feat: high-performance WebSocket streaming with GPU acceleration
HARI PREETHAMc736ef6 -
fix: restore turbo model and disable conditioning to fix loops
HARI PREETHAMf9ef6ba -
fix: upgrade to whisper large-v3 and increase context window
HARI PREETHAM53f5985 -
fix: multilingual support for Telugu script with large-v3-turbo
HARI PREETHAM0ba6048 -
feat: optimized distil-whisper and low-latency VAD
HARI PREETHAM016981b -
feat: enable NVIDIA T4 GPU support with CUDA 11.8
HARI PREETHAMb988fd5 -
feat: global refresh for HF deployment
HARI PREETHAMdddfb6b -
feat: complete backend redeploy with timestamp
HARI PREETHAM83bfc33 -
chore: prepare for huggingface deployment
HARI PREETHAM3b3890a
-
fix(backend): finalize HF environment and OpenMP settings
HARI PREETHAMb40434b -
fix(backend): apply user-corrected model names and ENV variables
HARI PREETHAMfbea35a -
fix(backend): whisper format and indictrans gated access on HF
HARI PREETHAMa165701 -
feat(backend): optimized ctranslate2 and granular health checks
HARI PREETHAMbde662a -
feat: optimized hybrid translation engine (IndicTrans2 + NLLB-CT2) and expanded language support
HARI PREETHAMd550119 -
feat: upgrade to whisper turbo and advanced VAD
HARI PREETHAM8aa0d82 -
fix: implement smart filtering with dictionary hints
HARI PREETHAM2a0ae22 -
fix: smart confidence-based transcription filters
HARI PREETHAM87001ac -
fix: implement hallucination blacklist and empty all prompts
HARI PREETHAM79e3985 -
fix: total cleanup of prompts and hallucination hardening
HARI PREETHAM92beb6e
-
fix: implement language-specific transcription guidance to fix Telugu spelling
HARI PREETHAMad0907c -
feat: upgrade Whisper model to 'large-v3' for maximum accuracy
HARI PREETHAM2c5a212 -
feat: upgrade Whisper model to 'medium' for higher accuracy
HARI PREETHAM3754495 -
deploy: load Whisper model in background for faster response
HARI PREETHAM414a93d -
deploy: fresh push from backend folder (source only)
HARI PREETHAM2ac469d