Bournemouth Forced Aligner (BFA)

⚠️ BFA 2.0 is not compatible with 1.x. The acoustic models are new: no CTC, no blank class, 5 ms frames, and a multilingual model family that is now the default. The alignment method has changed too. A segmental Viterbi, the same decoder the models were trained with, places each boundary where two phone posteriors cross, instead of snapping CTC spikes onto onsets. v1 checkpoints do not load, and the output and constructor arguments have changed: see the CHANGELOG. The last 1.x release is still on old the main commits (pip install "bournemouth-forced-aligner<2").

What it does

Given an audio file and its transcript, BFA returns a start and end time for every phoneme and word: a forced alignment. No pronunciation dictionary and no per-language training are needed.

Input:  butterfly.wav  +  "butterfly"

Output:
  b   →    0 ms –  52 ms  (confidence: 0.89)
  ə   →   52 ms – 168 ms  (confidence: 0.09)
  t   →  168 ms – 258 ms  (confidence: 0.05)
  ɝ   →  258 ms – 351 ms  (confidence: 0.01)
  f   →  351 ms – 498 ms  (confidence: 0.28)
  ɫ   →  498 ms – 572 ms  (confidence: 0.11)
  a   →  572 ms – 826 ms  (confidence: 0.94)
  ɪ   →  826 ms – 982 ms  (confidence: 0.25)

Install

pip install bournemouth-forced-aligner


# Install the latest version from github:
# pip install git+https://github.com/tabahi/bournemouth-forced-aligner.git

# Install the nightly release:
# pip install git+https://github.com/tabahi/bournemouth-forced-aligner.git@v2.0-nightly

# On install whatever version is currently on pypi

That is all you need for WAV, FLAC, OGG and MP3, on Linux, macOS (Intel and Apple silicon) and Windows. Optional extras:

pip install "bournemouth-forced-aligner[ja]"     # Japanese word segmentation (fugashi + unidic-lite)
pip install "bournemouth-forced-aligner[yue]"    # Cantonese word segmentation (pycantonese)
pip install "bournemouth-forced-aligner[ph66]"   # the espeak-ng phoneme set; also needs the espeak-ng system package
pip install "bournemouth-forced-aligner[audio]"  # other audio formats (m4a, aac, mp4); also needs FFmpeg

The first run downloads the aligner model (~55 MB) and, for the default phoneme set, the G2P model charsiu/g2p_multilingual_byT5_small_100 (~1.2 GB, once) into the Hugging Face cache. After that, everything runs offline.

Quick start

From the command line:

balign butterfly.wav "butterfly" --preset=en-us --textgrid=butterfly.TextGrid --mel-path=mel.png

This writes butterfly.vs.json next to the audio, a Praat TextGrid, and a plot:

From Python:

from bournemouth_aligner import BFA

aligner = BFA(preset="en-us")              # "de", "fr", "ru", "zh", "ja", ...
audio = aligner.load_audio("butterfly.wav")
result = aligner.process_sentence("butterfly", audio)

for ph in result["segments"][0]["phoneme_ts"]:
    print(ph["ipa_label"], round(ph["start_ms"]), round(ph["end_ms"]), round(ph["confidence"], 2))

aligner.convert_to_textgrid(result, output_file="butterfly.TextGrid")

To align a whole recording from a Whisper-style transcript ({"segments": [{"start", "end", "text"}]}):

result = aligner.process_srt_file("interview.srt.json", "interview.wav", ts_out_path="interview.vs.json")

balign interview.wav interview.srt.json interview.vs.json --preset=en-us

More: examples/, and the full API, CLI options and output format in the developer guide.

Languages and phoneme sets

BFA has two phoneme sets. Each is decoded by its own family of models:

BFA(preset="de")                         # std, mbfa-latin
BFA(preset="zh")                         # std, mbfa-cjk, with tone
BFA(preset="en-us", phoneme_set="ph66")  # espeak-ng, CUPE-3-en: the most accurate English model
BFA(preset="hi", phoneme_set="ph66")     # a language the std set has no model for yet

Every supported language, its code, and the details (tone, word segmentation, and why numbers must be spelled out) are on the language support page.

Models

Eight models, each a Hugging Face repository that downloads on first use. Each is a 13.5-14.9 M-parameter CNN that looks at 120 ms of audio at a time and outputs a phoneme posterior every 5 ms. The table, how they are trained, and how a boundary is read off them are on about the models.

Accuracy

Phoneme-onset F1 on full corpora (TIMIT 6,300 clips, Buckeye 255 recordings), compared with the Montreal Forced Aligner:

• BFA's onsets are more precise. The English model beats MFA within 10 ms and has a lower median error (7 vs 8 ms). At 25 ms MFA leads by 4 points, because BFA has more large misses.
• One model covers 36 languages, with no pronunciation dictionary. MFA needs a separate acoustic model and dictionary for each language.
• Buckeye favours MFA. Its reference boundaries were first placed by an HMM aligner with an ARPAbet dictionary, the same approach MFA uses. MFA scores higher on this noisy conversational corpus than on clean TIMIT.

The full analysis is on the evaluation page. A label-free test is coming soon. It stretches and compresses phones by known amounts, adds noise, and checks that the aligner measures the changes. It needs no hand labels, so it works in every language (see planned).

Settings worth knowing

FAQ

Do I need a pronunciation dictionary? No. The G2P model (or espeak-ng) writes the phonemes.

Numbers come out wrong. The std G2P reads dictionary words. Write "twenty five", not "25", or use phoneme_set="ph66", which expands numbers.

Confidence looks low. The models are trained not to saturate: a well-placed phone often reads below 0.5. Compare confidences within one model only. See confidence.

Words in the transcript are not in the audio. Every phone of the text is always placed, so extra words get squeezed into neighbouring audio. Check coverage_analysis["bad_alignment"].

Long recordings? Segments up to 120 s are aligned in one pass. Longer ones are walked in overlapping windows automatically. A transcript split into sentences still works best.

Contribute

• More language groups. Vietnamese, Greek/Armenian/Georgian/Korean, Arabic-script, Indic and Thai have token lists in the std set, but no trained model yet.
• ONNX / C++. bournemouth_aligner/cpp_onnx/ still targets the v1 model.

Open an issue to discuss. Start with the developer guide.