A short clip lives or dies at its edges. Cut 40 ms early and the first word loses its consonant. Cut late and you catch the breath before the next sentence.
So before cutting news into shorts, we need word timestamps we can trust. "Trust" has a specific meaning here: when we cut at a word boundary, does the cut land in silence?
The question
The public benchmark for this is FA-Bench, which scores aligners against hand-marked word boundaries in TIMIT and Buckeye. MFA leads it. But that is read speech and interviews, scored at a strict 20 ms. News has music beds, stings and interpreters, and our question is coarser and more practical: how often does a cut at a word boundary land inside a real pause?
What we ran
174 windows from real broadcasts, ASR transcripts as the text, on a CPU VM. Metric: percent of cuts that land inside a real pause. Higher is better.
| Aligner | Cuts in a pause |
|---|---|
| Chirp 3 (own transcript) | 85.3% |
| MFA 3.4.3 | 83.3% |
| Qwen3-ForcedAligner 0.6B | 81.0% |
| Meta MMS-FA | 80.0% |
| wav2vec2 CTC | 80.0% |
What we found
MFA 3.4.3 is the best open aligner: 83.3% of cuts land inside a real pause, with a 30 ms median error. Chirp 3 leads by 2.8 points (95% CI 0.9 to 5.2), mostly on pauses shorter than 0.3 s.
If you cut audio at word boundaries, the open option is close to the best closed one. The gap shows up on very short pauses.
This is stage 3 of NewsShorts-Bench. Method, windows and confidence intervals: https://xplormity.com/research/word-alignment-news-audio-2026-10
Top comments (0)