Spaces:
Running
Running
Commit History
-
Sync corrected docs into the Space image
Komalpreet KaurUnverified signatureDocstrings shipped in this image contradicted what it runs: inference.py said SepFormer 'delivers the audio quality we need for the live demo', speaker_encoder.py led with 'Pretrained ECAPA-TDNN', and sepformer_tse.py still justified itself with a '+1 dB ceiling' that no longer holds. The Dockerfile advertised superseded metrics and described SepFormer as a fallback that gets selected; requirements.txt described speechbrain as the speaker encoder. No behaviour change — the served models and settings are identical.47a4b2f
-
Fix residue: align estimate to mixture before subtracting
Komalpreet KaurUnverified signatureThe residue track was effectively a copy of the input (correlation +0.968, RMS higher than the mixture). SI-SDR is sign-invariant, so the separator often emits a phase-inverted waveform; subtracting that adds it back. Aligning by least squares recovers gain and sign together so extracted + residue reconstructs the input.046dbcd -
Deploy fully self-trained Vanta (separator + speaker encoder)
Komalpreet KaurUnverified signatureReplaces the pretrained ECAPA with our own trained speaker encoder, so no pretrained weights remain in the model. On 500 held-out mixtures the pair beats the previous deployment: SI-SDR +8.45 vs +7.93, PESQ 1.247 vs 1.182, STOI 0.751 vs 0.739. CPU inference is also ~6x faster without speechbrain's ECAPA. The separator was trained against this exact encoder, so both checkpoints ship together and must stay paired. Includes a decode fix for MP4/MOV whose moov atom follows the media data: those were piped to ffmpeg, which cannot rewind a pipe to find the index, and decoded to zero samples while exiting 0. Phone recordings often lack faststart, so this affected real uploads.1bd371e
-
Deploy from-scratch trained separator (val SI-SDR +7.93 dB)
Komalpreet KaurUnverified signatureSwitches the Space default from the pretrained SepFormer backbone to the model trained from scratch: 952 speakers (LibriSpeech + AMI conversational), WHAM real noise, real and simulated RIRs, turn-taking, recording-chain degradation. SepFormer still ships in the image and is available via VANTA_BACKEND=sepformer without a rebuild. Also fixes a device desync where speechbrain's ECAPA wrapper kept its own device attribute, which crashed CPU-only inference (as the Space runs).e1de6b8
-
sync: update server and scripts (model unchanged)
Komalpreet KaurUnverified signaturee51885a
-
Switch to SepFormer backend for +9.7 dB SI-SDR
Komalpreet KaurUnverified signature828f7dd
-
Initial deploy: Vanta TSE backend (FastAPI + trained checkpoint)
Komalpreet KaurUnverified signature32de4f6