Spaces:
Paused
Paused
Malina
Claude Fable 5.1
v0.90: scoped single-use tokens, hardened link fetch, 45-minute cap, whole-video digest, one classification call, no OCR
1d697a7 Download config.py from NextNeural/NINA: direct link, hf CLI and curl.
- Browser
- Download file 10 kB
-
https://huggingface.co/spaces/NextNeural/NINA/resolve/main/config.py
- Command line
-
hf download hf://spaces/NextNeural/NINA/config.py
-
curl -L -o config.py https://huggingface.co/spaces/NextNeural/NINA/resolve/main/config.py
10 kB
| import os | |
| import tempfile | |
| WHISPER_FAST = "base" | |
| WHISPER_PRECISE = "large-v3-turbo" | |
| SAMPLE_RATE = 16000 | |
| WHISPER_CPU_THREADS = int(os.getenv("WHISPER_CPU_THREADS", "3")) | |
| RETRANSCRIBE_MAX = 12 | |
| # Re-transcribe a sector with the precise model when the base model's confidence | |
| # (avg_logprob) drops below this. Loosened from -0.55 β -0.40 so it actually fires | |
| # on weak audio: the base model sits around -0.2β¦-0.5 on normal speech, so the old | |
| # gate almost never triggered. -0.40 catches genuinely-shaky sectors; the budget | |
| # (RETRANSCRIBE_MAX) and the per-sector cache keep it from running on everything. | |
| PRECISE_LOGPROB_GATE = -0.40 | |
| # ββ Whisper hallucination guards βββββββββββββββββββββββββββββββββββββββββββββββ | |
| # On noisy / musical / distorted audio, Whisper invents fluent-but-fake speech | |
| # (e.g. a concert recording β plausible nonsense). These decode-time thresholds | |
| # make it mark such audio as silence/low-confidence instead of fabricating, and | |
| # disable cross-segment conditioning so it can't spiral into repetition loops. | |
| WHISPER_NO_SPEECH_THRESHOLD = 0.6 | |
| WHISPER_LOGPROB_THRESHOLD = -1.0 | |
| WHISPER_COMPRESSION_RATIO_THRESHOLD = 2.4 | |
| WHISPER_CONDITION_ON_PREV = False | |
| # ββ ASR reliability β honest hedging ββββββββββββββββββββββββββββββββββββββββββ | |
| # A transcript whose mean confidence is poor (or which is mostly weak segments, | |
| # or mostly "no-speech") is flagged UNRELIABLE. Synthesis then tells the LLM the | |
| # audio is hard to make out and to say so plainly rather than treating garbled | |
| # fragments as real statements. | |
| ASR_RELIABILITY_LOGPROB_FLOOR = -0.85 # mean segment avg_logprob below β suspect | |
| ASR_WEAK_SEGMENT_LOGPROB = -0.85 # a single segment is "weak" below this | |
| ASR_UNRELIABLE_WEAK_FRACTION = 0.5 # β₯ this fraction weak β unreliable | |
| ASR_NO_SPEECH_CEILING = 0.55 # mean no_speech_prob above β unreliable | |
| CHUNK_SPLIT_THRESHOLD = 0.45 | |
| CHUNK_MIN_DURATION = 8.0 | |
| CHUNK_MAX_DURATION = 90.0 | |
| CHUNK_WINDOW = 5 | |
| BI_ENCODER_TOP_K = 40 | |
| CROSS_ENCODER_TOP_K = 10 | |
| RESULTS_TOP_N = 6 | |
| REGION_MIN_DURATION = 5.0 | |
| MERGE_GAP_SEC = 6.0 | |
| BI_SCORE_GLOBAL_FLOOR = 0.10 | |
| BI_SCORE_HYBRID_FLOOR = 0.28 | |
| GLOBAL_MAX_CHUNKS = 12 | |
| CLIP_MODEL = "clip-ViT-B-32" | |
| CLIP_MAX_FRAMES = 96 | |
| CLIP_MIN_FPS = 0.1 | |
| CLIP_MAX_FPS = 0.5 | |
| CLIP_FRAME_SIZE = 224 | |
| VISUAL_WEIGHT = 0.30 | |
| VISUAL_TAGS_TOP_K = 4 | |
| VISUAL_TAG_MIN_SIM = 0.20 | |
| VISUAL_CONCEPTS = [ | |
| "a person talking to the camera", "a group of people", "a presentation with slides", | |
| "a lecture or classroom", "an interview", "a news broadcast", | |
| "a screen recording of software", "source code in a text editor", | |
| "a webpage in a browser", "a chart or graph", "text on screen", | |
| "a product demonstration", "hands demonstrating something", | |
| "a cooking or food scene", "a kitchen", "an office", "a laboratory", | |
| "a factory or machinery", "a construction site", "an outdoor nature landscape", | |
| "a forest", "mountains", "a beach or ocean", "a city street", | |
| "buildings and architecture", "a car or vehicle driving", "an airplane", "a train", | |
| "a sports event", "people exercising", "a concert or live music", | |
| "a video game gameplay", "an animation or cartoon", "a medical or hospital scene", | |
| "a doctor or patient", "an animal", "a dog", "a cat", "a crowd of people", | |
| "a map", "a whiteboard with diagrams", "a scientific diagram", | |
| "a phone or mobile device", "a computer or laptop", "money or finance", | |
| "a meeting or conference room", "a stage with a speaker", | |
| "a dark scene at night", "a daytime scene", | |
| ] | |
| LLM_MODEL = "Qwen/Qwen2.5-72B-Instruct" # text-only fallback synthesis | |
| LLM_MAX_TOKENS = 1800 | |
| LLM_TEMPERATURE = 0.05 | |
| LLM_TIMEOUT = 120 | |
| # ββ Vision-Language synthesis (primary) βββββββββββββββββββββββββββββββββββββββ | |
| # NINA's answer engine is a VLM that actually SEES sampled keyframes + reads the | |
| # transcript, so it can read on-screen text, describe who/what is shown, and fuse | |
| # vision with speech β far more accurate than CLIP-tags + text-LLM. One call per | |
| # request. Falls back to the text LLM above if the VLM is unavailable. | |
| USE_VLM = True | |
| VLM_MODEL = "Qwen/Qwen3-VL-30B-A3B-Instruct" # live via novita/featherless; sharper | |
| # grammar/facts than the old 8B. Further | |
| # upgrade: "Qwen/Qwen3-VL-235B-A22B-Instruct" | |
| VLM_MAX_FRAMES = 10 # keyframes sent to the VLM (more = more image tokens) | |
| VLM_FRAME_MAX_PX = 768 # downscale each frame's long side (caps image tokens) | |
| # ββ Chat follow-ups βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ | |
| # Follow-up turns answer from the conversation + stored transcript, and before that | |
| # run two one-word checks (_route, _on_topic). All three used LLM_MODEL, the 72B | |
| # text model, and a single follow-up took 3 min 48 s on 2026-09-19: the route check | |
| # alone ran into the website's 30 s cap, and the answer streamed a 3.5k-char quiz in | |
| # nearly four minutes. In the same Space, minutes earlier, VLM_MODEL wrote the first | |
| # answer in 8 s. Measured directly it generates ~4.5x faster (~450 vs ~100 chars/s), | |
| # and it is the model that already writes every analysis, so the voice stays the same. | |
| CHAT_MODEL = VLM_MODEL | |
| # The two checks have safe defaults (route -> answer, on_topic -> on), so a slow | |
| # provider must not be allowed to hold a turn hostage. Past this, take the default. | |
| CHAT_CHECK_TIMEOUT = 10 | |
| # Server-side caps on what one chat turn may carry (the website caps too, but a token | |
| # holder can call the Space directly). | |
| CHAT_MAX_TRANSCRIPT_CHARS = 60000 | |
| CHAT_MAX_MESSAGE_CHARS = 4000 | |
| CHAT_MAX_PROMPT_CHARS = 2000 | |
| # ββ Pasted links ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ | |
| FETCH_DEADLINE_SEC = 240 # the website gives up at 4 minutes anyway | |
| FETCH_MAX_BYTES = 400 * 1024 * 1024 | |
| # ββ Upload limits & server-side compression βββββββββββββββββββββββββββββββββββ | |
| # The website rejects > MAX_UPLOAD_MB before upload; we re-check here because the | |
| # client check is bypassable. Files over COMPRESS_TRIGGER_MB are transcoded down | |
| # to an aggressive 720p / low-bitrate MP4 before decoding β NINA only needs the | |
| # audio plus a handful of sampled frames, so quality loss is irrelevant and it | |
| # makes frame extraction + decode noticeably faster on big uploads. | |
| MAX_UPLOAD_MB = 500 | |
| # Longer than this is refused before any decoding: the waveform of 45 minutes alone is | |
| # ~170 MB of float32, and nothing above it finishes inside the website's turn budget. | |
| MAX_VIDEO_SECONDS = 45 * 60 | |
| COMPRESS_TRIGGER_MB = 20 | |
| COMPRESS_TARGET_HEIGHT = 720 | |
| COMPRESS_VIDEO_BITRATE = "1200k" | |
| COMPRESS_AUDIO_BITRATE = "96k" | |
| # ββ Vocal isolation (torchaudio Hybrid Demucs) ββββββββββββββββββββββββββββββββ | |
| # Pull the speaking voice out from under music/noise before re-transcribing, but | |
| # only when the first ASR pass looks unreliable (musical/distorted) β it's CPU-heavy. | |
| # Uses torchaudio's bundled Hybrid Demucs (no separate `demucs` package β no torch | |
| # version conflict). Best-effort: silently skipped if torchaudio is unavailable. | |
| USE_VOCAL_ISOLATION = True | |
| DEMUCS_BUNDLE = "HDEMUCS_HIGH_MUSDB_PLUS" | |
| ISOLATION_MAX_SECONDS = 480 # don't isolate very long audio (CPU/memory) | |
| ISOLATION_CHUNK_SEC = 12.0 # process in chunks to bound memory | |
| # ββ Subtitle fallback ladder ββββββββββββββββββββββββββββββββββββββββββββββββββ | |
| # When the ASR transcript is unreliable (garbled/musical), prefer a clean human | |
| # subtitle track if the file carries one. Only text-based subtitle codecs can be | |
| # read directly; bitmap subs (PGS/VobSub) would need OCR (separate, heavier step). | |
| USE_SUBTITLES = True | |
| SUBTITLE_TEXT_CODECS = frozenset({ | |
| "subrip", "srt", "ass", "ssa", "mov_text", "webvtt", "vtt", "text", "eia_608", | |
| }) | |
| # OCR of burned-in captions. Disabled now that the VLM reads on-screen text | |
| # natively (and far better) β kept here so it can be re-enabled if ever needed. | |
| USE_OCR = False | |
| OCR_FPS = 1.0 # sample ~1 frame/sec | |
| OCR_MAX_FRAMES = 90 | |
| OCR_FRAME_W = 1280 | |
| OCR_FRAME_H = 720 | |
| OCR_BOTTOM_FRACTION = 0.34 # crop the bottom third (where captions sit) | |
| VISUAL_ONSCREEN_TEXT_SIM = 0.22 # CLIP gate: frames must look like they contain text | |
| # ββ Speaker diarization (pyannote) + name attribution βββββββββββββββββββββββββ | |
| # Identifies who-spoke-when; synthesis then attributes lines and maps a speaker to | |
| # a name when they introduce themselves. Gated models β the Space's HF_TOKEN must | |
| # have accepted pyannote/speaker-diarization-3.1 + pyannote/segmentation-3.0. | |
| # Best-effort: skipped if pyannote is unavailable or the token lacks access. | |
| USE_DIARIZATION = True | |
| DIARIZATION_MODEL = "pyannote/speaker-diarization-3.1" | |
| DIARIZATION_MAX_SECONDS = 1800 | |
| CACHE_DIR = os.path.join(tempfile.gettempdir(), "nina_cache") | |
| DISK_CACHE = True | |
| CACHE_IDLE_TTL = 1200 | |
| CACHE_DISK_TTL = 1800 | |
| DEBUG = False | |
| SIGNING_SECRET = os.getenv("NINA_SIGNING_SECRET", "") | |
| SPACE_TAG = "nina" | |