Mixed audio · English + Mandarin · Phased evaluation

TS-ASR Benchmark

A phased benchmark for acoustic event grounding, speaker-aware transcription, and target-voice extraction from mixed audio.

Explore text output Explore audio output
Phase 0 is under development · Phase 1 and Phase 2 demos are reviewable
3benchmark phases
10,327active questions
7source datasets
10speaker clue types

Benchmark structure

Three phases, three explicit outputs

Each phase has its own task definition, output contract, evaluation metrics, and demo surface.

00 Acoustic Event GroundingLocate one described audible event in mixed audio. Time interval In development 01 Speaker-Aware TranscriptionTranscribe every speaker or only the speaker resolved from a clue. Timestamped text 21 demos 02 Target Voice ExtractionReturn an isolated target waveform under one of two audio contracts. Target audio 7 demos

Phase 1 · Text output

Speaker-Aware Transcription

Phase 1 separates complete room transcription from clue-conditioned target transcription. Phase 1.3 extends the same text task with stable speaker-timbre descriptions.

InputMixed speech, with an optional speaker clue
OutputSpeaker-attributed timestamped text
EvaluationcpWER · DER · temporal validity
View

0 of 0 demos

Preparing Phase 1 demos...

Phase 2 · Audio output

Target Voice Extraction

Phase 2 reuses the speaker-clue vocabulary but changes the model target from text to audio. Timeline-preserved and speech-only outputs are evaluated separately.

InputFull mixed scene with a speaker clue
OutputTimeline-preserved or speech-only target waveform
EvaluationSI-SDR · SI-SDRi · length error
View

0 of 0 demos

Preparing Phase 2 demos...

00
Phase 0In development

Acoustic Event Grounding

Given mixed audio and a natural-language description of an audible event, return one start/end interval. The task schema, candidate boundaries, and evaluation protocol are still under review.

Open Phase 0 preview

Evaluation contracts

Outputs are scored within their phase

No metric is shared across incompatible text, timeline, and waveform outputs.

PhaseReference outputPrimary metricsStatus
Phase 0One acoustic event intervalTemporal-F1 · Boundary-Hit@1sDraft
Phase 1Timestamped speaker turnscpWER · DERActive
Phase 2Isolated target waveformSI-SDR · SI-SDRi · durationPilot
1

Deterministic selectionEvery public demo is selected explicitly in versioned configuration.

2

Audio verificationReferenced audio, target stems, and output contracts are validated before export.

3

Public-safe metadataThe static bundle contains no private filesystem paths or raw source data.