Bournemouth Forced Aligner (BFA)

• BFA is a forced alignment system that maps known phoneme transcripts onto speech audio with explicit modelling of inter-phoneme gaps and silence.
• It integrates a Contextless Universal Phoneme Encoder (CUPE) with a CTC-based dynamic decoder and hierarchical segmentation to support multilingual and real‐time applications.
• Evaluations show BFA achieves competitive recall with significantly faster processing times than traditional aligners like MFA, enabling interactive speech applications.

Searching arXiv for Bournemouth Forced Aligner and closely related forced-alignment work. Bournemouth Forced Aligner (BFA) is a forced alignment system for mapping a known phoneme transcript onto speech audio with fine temporal boundaries. It combines a Contextless Universal Phoneme Encoder (CUPE) with a connectionist temporal classification (CTC)-based decoder, introduces explicit modelling of inter-phoneme gaps and silences, and uses hierarchical decoding strategies for long or segmented utterances. In the reported evaluation, BFA is designed to be fast enough for real-time use, multilingual or language-agnostic, silence-aware, and capable of predicting both onset and offset boundaries for phonemes rather than a single boundary point (Rehman et al., 27 Sep 2025).

1. Motivation and problem setting

BFA is presented as a response to several limitations attributed to existing forced aligners. Traditional tools such as Montreal Forced Aligner (MFA) are described as accurate but computationally expensive. The reported timings emphasize this point: MFA can take about 45 minutes per 535-second Buckeye clip and about 1 minute per 3.1-second TIMIT clip, which makes large-scale or interactive alignment difficult. The same summary also states that modern ASR-based systems such as WhisperX are faster than some classical pipelines, but in pure alignment tasks they often underperform MFA (Rehman et al., 27 Sep 2025).

A second motivation is representational rather than computational. The paper argues that many aligners usually ignore silences or gaps between phonemes. BFA therefore introduces explicit modelling of inter-phoneme gaps and silence, with the stated aim of supporting richer temporal structure and prosody-sensitive applications. A third motivation is multilinguality: the system is intended to work across languages without being tightly coupled to a single pronunciation dictionary or language-specific acoustic model. This places BFA within a line of research that treats forced alignment not merely as a transcription utility but as a temporal-structure estimation problem.

The resulting design target is therefore not only faster alignment, but a broader output object. In BFA, phonemes are treated as intervals with potentially nonzero gaps between them, rather than as a strictly back-to-back sequence of single-boundary events. This difference is central to both the claimed strengths and the reported metric trade-offs.

2. Core architecture and decoding formalism

The paper describes BFA as having three main parts: a CUPE-based acoustic classifier, a multilingual phonemization module, and a CTC-style temporal decoder (Rehman et al., 27 Sep 2025).

CUPE is the acoustic front end. In this formulation, it is adapted for forced alignment and trained to classify speech frames into phoneme-related labels without relying on left or right phonetic context. The model is implemented with two prediction heads: a 67-class phoneme prediction head and a 17-class phoneme-group prediction head, both including blank tokens for CTC training. The stated purpose is to support alignment at two granularities: fine-grained phoneme alignment and coarse phoneme-group alignment. Different model variants are trained on different corpora: BFA_en on LibriSpeech, BFA_eu on Multilingual LibriSpeech (MLS) across 7 languages excluding English, and BFA_world on MSWC, covering 35 languages excluding English.

The multilingual phonemization module uses espeak-ng to convert text into phonemes and then maps those phonemes to the paper’s internally defined phoneme set. The summary states that this reduces dependence on language-specific pronunciation dictionaries and manually engineered phoneme inventories, while using a more data-driven universal phoneme mapping optimized for Indo-European coverage in the training data. It also notes that broader phoneme groups can be formed for coarse-grained alignment.

Forced alignment itself is performed with a CTC-style dynamic program. Given a target phoneme sequence of length $S$, the decoder constructs the path

$$\text{path} = [\text{blank}, p_1, \text{blank}, p_2, \ldots, \text{blank}, p_S, \text{blank}]$$

so that blank states allow variable durations and monotonic alignment. The decoding recursion is reported as

$$\alpha_t(s) = \max \begin{cases} \alpha_{t-1}(s) + \log P(o_t \mid s) & \text{(duration extension)} \ \alpha_{t-1}(s-1) + \log P(o_t \mid s) & \text{(state transition)} \ \alpha_{t-1}(s-2) + \log P(o_t \mid s) & \text{(blank skip)} \end{cases}$$

where the three cases correspond to staying in the same state, moving to the next state, or skipping a blank state. The blank-skip transition is constrained to avoid illegal consecutive identical phoneme transitions, thereby preserving monotonic and valid alignments.

3. Boundary representation, silence modelling, and post-processing

A major contribution of BFA is its treatment of silence and inter-phoneme gaps. The paper states that BFA introduces explicit modelling of inter-phoneme gaps and silences, meaning that after one phoneme ends, the system can predict a time interval before the next phoneme starts. Conventional aligners are described as usually placing phonemes back-to-back and annotating only one boundary, whereas BFA predicts both onset and offset boundaries, so each phoneme is represented as an interval rather than as a single time point (Rehman et al., 27 Sep 2025).

This interval-based representation enables explicit modelling of phoneme duration, pauses between phonemes, prosodic structure, and inter-phoneme gaps. The paper further notes that many predicted offsets occur before the next phoneme onset, creating an interval of silence or gap. Roughly 30–40% of phonemes show such gaps in the reported datasets. A direct consequence is a metric asymmetry: because BFA may generate roughly twice as many boundary points as onset-only systems, conventional precision metrics can appear lower even when the temporal structure is richer. For this reason, the paper explicitly notes that precision at 20 ms is calculated using only onset boundaries to keep comparison with conventional systems fair.

For utterances with detected silence regions, BFA uses a divide-and-conquer hierarchical decoding strategy: it detects continuous speech regions, aligns each region independently, and then combines the results. The stated motivation is to reduce decoding complexity and improve handling of long utterances with pauses.

The system also includes several lightweight correction mechanisms. One is completeness enforcement, a post-processing step ensuring 100% target phoneme coverage: if a target phoneme is missing in decoding, it is inserted at the frame with maximum probability. Another is probability boosting, implemented with a logarithmic boosting factor

$\beta = 5.0$

to increase the probability mass of expected phonemes, especially in cross-lingual or universal settings where some target phones may be underrepresented. A minimum probability threshold

$\epsilon = 10^{-8}$

is also applied so that target phonemes are not completely eliminated during decoding. The reported ablations indicate that probability boosting helps especially in the universal model, and that 100% coverage enforcement slightly improves recall for the universal model without significantly harming boundary distance accuracy (Rehman et al., 27 Sep 2025).

4. Evaluation protocol and quantitative results

The reported evaluation uses two datasets. TIMIT contains 6300 clean read utterances with average duration 3.1 seconds and is used for standard phoneme alignment evaluation. Buckeye consists of spontaneous conversational speech from 40 speakers, with average clip duration 535 seconds, and is described as more difficult because of disfluencies and natural speech variation. The main baseline is Montreal Forced Aligner (MFA). Metrics include recall and precision at 20 ms, 40 ms, and 60 ms tolerance levels, along with boundary distance errors, deletions, insertions, and statistics on inter-phoneme gaps (Rehman et al., 27 Sep 2025).

On TIMIT, recall is reported as 71.9 / 81.2 / 82.8 for MFA, 71.4 / 84.6 / 87.9 for BFA_en, 71.0 / 84.7 / 88.1 for BFA_eu, and 60.9 / 73.2 / 77.4 for BFA_world. At relaxed tolerances, BFA_en and BFA_eu outperform MFA in recall. Precision at 20 ms is reported as 81.2 for MFA, 55.6 for BFA_en, 55.2 for BFA_eu, and 50.9 for BFA_world; the paper stresses that this is partly a metric mismatch because BFA predicts richer boundary structure.

On Buckeye, recall is reported as 58.1 / 70.8 / 74.8 for MFA, 63.7 / 73.2 / 76.3 for BFA_en, 60.5 / 70.6 / 73.5 for BFA_eu, and 58.6 / 67.7 / 71.0 for BFA_world. Here BFA_en exceeds MFA on recall across all tolerance levels. Precision at 20 ms is 58.5 for MFA, 48.6 for BFA_en, 47.9 for BFA_eu, and 44.8 for BFA_world, again reflecting the onset/offset representation.

Boundary-distance statistics are also reported. On TIMIT, the known-to-aligned mean and aligned-to-known mean are 13.2 ms and 11.8 ms for MFA, 14.1 ms and 20.0 ms for BFA_en, 14.4 ms and 20.1 ms for BFA_eu, and 16.4 ms and 21.0 ms for BFA_world. On Buckeye, the corresponding values are 16.8 ms and 17.9 ms for MFA, 13.6 ms and 20.9 ms for BFA_en, 13.7 ms and 20.9 ms for BFA_eu, and 14.4 ms and 21.4 ms for BFA_world. The summary interprets these results as showing that BFA is competitive in boundary proximity, while offset-side errors are somewhat larger because of interval and gap modelling.

Inter-phoneme gap analysis further differentiates the systems. On TIMIT, about 33–35% of phonemes have gaps under BFA; on Buckeye, about 32–34% do. The median gap for BFA is around 39–41 ms. By contrast, MFA is described as effectively having almost no such gap modelling in the reported statistics.

5. Throughput, real-time operation, and application scope

Speed is one of the paper’s central claims. The reported processing times are approximately 45 min per clip for MFA versus ~60 s per clip for BFA on Buckeye, and ~1 min per clip for MFA versus ~0.25 s per clip for BFA on TIMIT. The paper therefore reports BFA as 45–240× faster than MFA. In terms of real-time factor, MFA is reported as 52–194× slower than real time, while BFA is reported at 0.05–0.1× real-time factor, i.e., faster than real time (Rehman et al., 27 Sep 2025).

The summary attributes this throughput to several design choices: contextless phoneme encoding, CTC-style monotonic decoding, hierarchical segmentation at silence regions, universal phoneme mapping, and post-processing instead of complex decoding. Taken together, these design decisions are presented as the reason BFA can support faster-than-real-time alignment while maintaining competitive recall relative to MFA at relaxed tolerance levels.

The reported application implications are correspondingly oriented toward latency-sensitive use. The paper argues that BFA opens up forced alignment to interactive speech applications, real-time text-to-speech pipelines, live speech editing, multilingual speech processing, audio-visual synchronization, and prosody analysis. Because the output includes explicit onset, offset, and gap information, these use cases are not limited to transcript anchoring; they also involve temporal structure that is useful when pauses and sub-utterance timing matter.

6. Relation to multilingual and low-resource forced alignment, limitations, and bibliographic issues

BFA is explicitly framed as multilingual or language-agnostic, and the model variants BFA_eu and BFA_world were trained on multilingual corpora. However, the reported evaluation is only on English test sets, namely TIMIT and Buckeye. The paper itself identifies this as a limitation and states that future work should test on tonal languages and broader multilingual settings. It also notes that conventional precision can appear lower because BFA predicts extra offset boundaries and gaps, and suggests future work on adaptive boundary prediction to tune precision–completeness trade-offs depending on application needs (Rehman et al., 27 Sep 2025).

A relevant contextual comparison comes from work on low-resource related languages using Montreal Forced Aligner rather than BFA. That study compares training from scratch with adaptation from a large pretrained English model on several Australian Indigenous languages and concludes that, in those low-resource settings, a large pretrained English model adapted to the target or related-language data is usually the strongest choice, especially for a language not seen in training. It also reports a notable 29% improvement from multilingual scratch training on the unseen-language test, while still finding English-based models more robust overall (Tosolini et al., 9 Apr 2025). This does not constitute a direct result about BFA, but it suggests that multilingual alignment systems are likely to benefit from strong pretrained acoustic representations and careful adaptation regimes when moving beyond well-resourced evaluation conditions.

The same low-resource study also emphasizes an important evaluation caveat: phone boundaries are inherently fuzzy and human annotators disagree, so “human annotated” boundaries should not be treated as perfectly exact (Tosolini et al., 9 Apr 2025). A plausible implication is that some of BFA’s apparent precision deficit under onset-only metrics should be interpreted in the context of both representational mismatch and the intrinsic uncertainty of boundary annotation.

A separate bibliographic issue concerns arXiv record (Lou et al., 2024). Although its title and abstract describe an “Aligner-Guided Training Paradigm” for text-to-speech, the supplied details state that the underlying document is an IEEE conference LaTeX template containing only formatting guidance and placeholder text, with no substantive content about forced alignment, BFA, MFA, or TTS (Lou et al., 2024). This is relevant because it cautions against treating all ostensibly alignment-related metadata as technically valid source material.

In sum, BFA is characterized in the literature as a CUPE + CTC forced aligner whose distinctive features are explicit silence/gap modelling, onset-and-offset interval prediction, and hierarchical decoding. Its reported empirical profile is a combination of competitive recall, especially at relaxed tolerances, and substantially higher throughput than MFA, with the trade-off that conventional boundary precision metrics do not fully reflect its richer temporal representation (Rehman et al., 27 Sep 2025).