We show that forced aligners degrade on spontaneous, noisy real-world speech, and that fillers, laughter, and disfluencies corrupt word boundaries, impacting many downstream applications. We revisit the classical GMM-HMM aligner with upsampled WavLM features instead of MFCCs, explicit phone units for fillers and non-lexical vocal events trained on verbatim-annotated spontaneous speech, and a discriminative projection fitted on clean and noise-corrupted feature pairs to improve robustness. The pipeline is self-bootstrapped from 149h of curated audio without manual or external boundary targets. On conversational (Buckeye) and disfluent (FluencyBank) speech it reaches word-boundary MAEs of 20.6 and 76.9ms, 19% and 17% lower than the strongest of ten baselines. Catastrophic misalignments (>200ms) are 2.7 times rarer than Montreal Forced Aligners (MFA) best model even after widening MFA's search beam, which we show removes a large pruning artifact from its conversational error; event units halve the error next to laughter, while being meaningfully more robust than existing aligners.
Figures & tables
System
Front-end
Boundaries
Lbl.-free
Events
Noise
MFA
MFCC
HMM
✓
×
×
CTC family
w2v2
CTC peaks
✓
×
∼
Charsiu
w2v2
frame
×
×
×
Qwen3-FA
LLM
tokens
×
×
∼
CrisperWhisper
Whisper
cross-attn.
×
×
∼
CrisperWhisper 2.0
Whisper
cross-attn.
×
✓
∼
Table 1: Capability grid. Lbl.-free = trained without manual or external boundary labels; ∼ = partial robustness inherited from augmentation or large-scale pretraining.
Source
Hours
Utts
Contents
ICSI meetings
39.2
24.7k
multi-party meetings
NSC part 6
24.5
4.0k
conversational, accented
LibriSpeech+events
21.2
4.9k
audiobooks, event overlay
Common Voice 14
20.1
11.5k
crowd-read prompts
AMI meetings
19.9
27.3k
multi-party meetings
VCTK
15.0
16.2k
studio-read, many speakers
Table 2: Training corpora. Transcripts follow a common verbatim convention; event tags are included where annotated or synthetically inserted.
Buckeye (conversational)
TIMIT (read)
FluencyBank (disfluent)
System
MAE
≤ 25
≤ 50
≤ 100
≤ 200
MAE
≤ 25
≤ 50
≤ 100
≤ 200
MAE
≤ 25
≤ 50
≤ 100
≤ 200
Competing systems
Qwen3-FA-0.6B
39.5
52.8
85.9
96.6
98.6
30.0
51.4
83.6
96.8
99.9
109.5
35.1
61.4
81.9
89.8
MMS-FA
53.0
57.5
81.3
92.8
97.8
37.2
51.5
82.6
97.5
99.9
158.8
35.7
56.8
74.1
86.4
CrisperWhisper 2.0
38.7
54.2
78.7
91.7
97.6
31.2
58.5
82.6
92.5
98.8
93.2
39.8
62.2
80.3
90.6
CTC-Segmentation
81.0
41.5
69.5
89.3
97.4
49.3
35.3
64.9
90.9
99.3
206.6
22.5
46.0
68.6
85.0
Table 3: Word-boundary MAE (ms, ↓ ) and offset threshold accuracies (%, ↑ ) on Buckeye (conversational), TIMIT (read), and FluencyBank (disfluent). All WavLM arms include the LDA projection of Sec. 3 ; the final row replaces it with the noise-paired LDA, all other components unchanged.
Figure 1: Complementary CDF (log–log) of the per-word boundary error δ on Buckeye, all aligners at converged (wide-beam) search: each curve gives the share of words misaligned by more than t ms. Ours is the only system whose tail falls below 1% before 100 ms.
System
filler-adj.
laugh-adj.
cutoff-adj.
Competing systems
Qwen3-FA-0.6B
56.4
180.4
70.8
MMS-FA
80.4
154.2
72.1
CrisperWhisper 2.0
66.0
82.3
69.6
CTC-Segmentation
170.0
227.6
101.2
Charsiu
162.3
269.5
91.3
Table 4: Buckeye MAE (ms, ↓ ) on regular words adjacent to a filler ( n =2,656), a laugh ( n =722), or a word cutoff ( n =1,369).
MUSAN noise
Gaussian noise
System
20
10
5
0
20
10
5
0
Competing systems
Qwen3-FA-0.6B
41
54
79
154
43
85
154
339
MMS-FA
53
55
58
67
52
55
61
99
CrisperWhisper 2.0
39
42
48
56
40
51
67
103
CTC-Segmentation
80
82
94
136
79
91
140
410
Table 5: Buckeye word-boundary MAE (ms, ↓ ) under additive noise at SNRs of 20 to 0 dB (column headers): held-out MUSAN noise clips and unseen Gaussian noise (training augmentation: MUSAN music only).