agwer 0.4.2

Agentic WER: WER/MER/WIL/WIP/CER plus RIR, HER, and n-best oracles for evaluating ASR error correctors and voice agents. Self-contained, single dependency (RapidFuzz).

AgWER: Agent-oriented Word Error Rate

agwer is a simple and efficient Python package to evaluate speech recognition and voice agents. A 40 KB wheel, and it imports in under 20 ms with fast CPU and native Apple Silicon on M-series supports.

agwer supports the classic ASR similarity measures and the agentic ones:

1. word error rate (WER) and sentence error rate (SER)
2. character error rate (CER)
3. recoverable information ratio (RIR)
4. harmful edit rate (HER)
5. named entity F1 score (NF1)
6. word hallucination rate (WHR)

The agentic measures (3 and 4) evaluate systems that read n-best hypotheses and decide when to edit and when to abstain, such as LLM error correctors and dictation agents. The metric guide explains what question each measure answers, and the example session walks through a complete evaluation from install to CLI.

Installation

With uv:

uv add agwer             # as a project dependency
uv pip install agwer     # into the active environment

Or with pip (Python >= 3.9):

pip install agwer

The test suite ships inside the package, so any install can verify itself:

pip install "agwer[test]"
python -m pytest --pyargs agwer     # 68 tests, a few seconds

Usage

The simplest use case is computing the word error rate of a dictated utterance:

import agwer

ref  = "please schedule the quarterly budget review for tuesday march twenty first at nine thirty and invite the design team"
asr_decoded = "please schedule the quarterly budget review for tuesday march twenty first at nine thirty and invite the desire team"

agwer.wer(ref, asr_decoded)   # 0.0526, one broken word in nineteen

All measures accept a single string or a list of strings. Lists are pooled corpus-level (total errors over total reference words), and mer, wil, wip, cer work the same way:

import agwer

refs = ["send the revised contract to the legal team before the board meeting on friday afternoon",
        "remind me to pick up the prescription from the pharmacy after the dentist appointment"]
hyps = ["send the revised contract to the legal team before the bored meeting on friday afternoon",
        "remind me to pick up the prescription from the pharmacy after the dentist appointment"]

agwer.wer(refs, hyps)     # 0.0345, corpus WER
agwer.cer(refs, hyps)     # 0.0116, corpus CER

Evaluating a corrector or voice agent

With n-best input, one call computes everything. Real data end to end: three real Whisper 5-best decodes from the MIT-licensed HyPoradise benchmark (WSJ), each with the verbatim output of a real LLM corrector. The three utterances show the three things a corrector can do: restore a word from another hypothesis, fix a reading that every hypothesis got wrong, and break a name that was already right.

refs = [
    "quote obviously we were disappointed we did not get a larger award",
    "he retired as a partner in nineteen eighty three and as counsel in nineteen eighty six",
    "a senior painewebber official said the firm hopes the job can be cut through attrition since the turnover in such positions tends to be high",
]

year = ("he retired as a partner in one thousand, nine hundred and eighty-three "
        "and as counsel in one thousand, nine hundred and eighty-six")
firm = ("official said the firm hopes the job can be cut through attrition "
        "since the turnover in such positions tends to be high")
nbest = [   # real Whisper 5-best lists; nbest[i][0] is the 1-best
    ["obviously we were disappointed if we did not get a larger award",
     "quote obviously we were disappointed we did not get a larger award",
     "obviously we were disappointed if we did not get a larger award",
     "quote obviously we were disappointed we did not get a larger award",
     "quote obviously we were disappointed we did not get a larger award"],
    [year] * 5,   # all five hypotheses agree on the wrong year reading
    [f"a senior {x} {firm}" for x in
     ["painewebber", "payne weber", "payne weber", "paine webber", "pain weber"]],
]

lm_corrected = [   # the LLM corrector's verbatim outputs
    "quote obviously we were disappointed we did not get a larger award",
    "he retired as a partner in nineteen eighty three and as counsel in nineteen eighty six",
    "a senior paine webber official said the firm hopes the job can be cut through attrition since the turnover in such positions tends to be high",
]

out = agwer.evaluate(refs, lm_corrected, nbest=nbest)
out.wer_1best            # 0.2264  the raw ASR
out.wer_oracle           # 0.1887  o_nb: the best any reranker could reach
out.wer_compositional    # 0.0377  o_cp: recombining n-best tokens cannot write "nineteen"
out.wer_corrected        # 0.0377
out.rir                  # 5.0     five times the reranking headroom: generative correction
out.her                  # 0.3333  one of the three consequential edits broke a correct name

Real decodes also show why reranking alone cannot save a voice agent. In another HyPoradise utterance, the command "leaving after noon" comes back as "leaving afternoon" in all five hypotheses. The query's meaning flips, every reranker is helpless, and only a corrector (and agwer's oracles) can see it.

When the corrector beats every hypothesis (ρ > 1)

Look at the second utterance above. All five hypotheses read the year the same wrong way, so the reranking oracle equals the 1-best and offers zero headroom. Recombining n-best tokens cannot help either, because the word "nineteen" appears nowhere in the list. The corrector still writes "nineteen eighty three", the dictation convention it knows from context. Supplying truth that exists in no hypothesis is generative correction. It is why ρ can exceed 1, and it is exactly what plain WER, reranking metrics, and oracle bounds cannot see. RIR was built to measure it.

Dictating to a coding agent has the same structure: package names and code terms (uv, pytest, agwer) are exactly the tokens ASR breaks and exactly the ones the agent can restore from context.

On the full 30-utterance WSJ set these corrections come from, the corrector scores ρ = 2.0 with HER = 0.29: it recovers twice the n-best headroom, and roughly two of every seven consequential edits still break something. That pair of numbers, not WER alone, is the act/abstain profile agwer exists to report.

Entity F1 and Word Hallucination Rate

Two failure modes that plain WER cannot name. First, entity errors hide inside acceptable WER: one substituted amount in a 21-word transfer request is a comfortable 4.8% WER and a broken transaction. entity_f1 scores an explicit token subset with the same alignment engine as WER:

m = agwer.entity_f1(ref, hyp, entities={"five", "hundred"})
m["recall"], m["f1"]                      # 0.5, the layer WER hid
agwer.entity_f1(refs, hyps, predicate=agwer.numeric_tokens)   # digits and spelled numbers

Second, LLM correctors make text up, and the two mechanisms differ. A corrector can invent a word the ASR never produced, or it can loop autoregressively and repeat words it really heard (one system in arXiv:2408.16180 repeats a real segment eleven times, so a vocabulary check sees nothing wrong). word_hallucination_rate uses occurrence-bounded attribution to catch both, works with a single 1-best or a full n-best list, and never penalizes copying an ASR error or a correct generative recovery:

m = agwer.word_hallucination_rate(refs, outputs, onebest_or_nbest)
m["whr"]                          # hallucinated tokens / output tokens
m["novel_hallucinations"]         # invented words
m["repetition_hallucinations"]    # autoregressive loops over heard words
m["passed_through_errors"]        # copied ASR errors: not hallucination
m["generative_tokens"]            # correct words no hypothesis contained

HER comes in two granularities. her_granularity="utterance" (the default) is the accounting behind the paper's reported values, and "token" follows the formal per-edit definition. A sentence where the corrector fixes one token and breaks another is neutral at utterance granularity but helpful=1, harmful=1 at token granularity, so report which one you used.

Normalization

Normalization is the main reason WER numbers are incomparable across papers. Every entry point takes normalize= (any Callable[[str], str]), and agwer ships the standards.

A dictation agent that says amounts out loud looks 87% wrong against the written form, until you normalize:

ref = "the invoice total came to $1,250.75 after the 15% discount was applied on march 3rd"
hyp = ("the invoice total came to one thousand two hundred fifty dollars "
       "and seventy five cents after the fifteen percent discount was applied on march third")

agwer.wer(ref, hyp)                                          # 0.867 (!)
agwer.wer(ref, hyp, normalize=agwer.EnglishTextNormalizer())  # 0.0

normalizer	what it does
None (the general measures' default)	scores strings exactly as given
agwer.default_normalize (the agentic default)	conservative: lowercase, keep apostrophes, strip other punctuation
BasicTextNormalizer()	language-agnostic: symbols, brackets, optional diacritic folding
EnglishTextNormalizer()	the Whisper English normalizer: spelled numbers to digits, currency, contractions, British to American spelling; cached=True adds an LRU for agent loops

The Whisper normalizers have their behavior pinned byte-identical to the original by golden tests. Report which normalizer you used; it is part of the metric.

CLI

agwer results.jsonl                 # {"reference","corrected","nbest"} per line
agwer results.jsonl --json --her-granularity token

Performance

The hot path is batched RapidFuzz, and every aggregate agwer computes is count-additive, so large corpora parallelize exactly. Results are identical for any worker count: evaluate(..., workers=8).

Typical performance on Apple Silicon (M-series, 14 cores) for the full agentic evaluation (three WERs, both oracles, RIR, and HER) of 5-best corpora:

scenario	time
10k utterances (single-threaded)	~0.2 s
100k utterances (single-threaded)	2.6 s
100k utterances (8 workers)	0.48 s (5.5×)
1M utterances (single-threaded)	27.9 s
1M utterances (8 workers)	5.2 s (5.4×)

Workers pay process startup, so they win from roughly 100k utterances up. On macOS and Windows, call from a if __name__ == "__main__" guard, as with any multiprocessing. Reproduce on your machine:

python -m agwer.bench --workers 8

Apple Silicon

agwer is native on Apple Silicon, with no separate install: pip and uv select the arm64 wheel automatically, and RapidFuzz ships compiled arm64-darwin extensions, so the C++ edit-distance core runs natively on M-series machines. workers= then scales the whole pipeline across performance cores (see the table above).

• Planned next is an optional agwer[mlx] extra for embedding-based semantic metrics on the Apple GPU via MLX. i.e., Semantic inference is the one place extra hardware genuinely helps; edit distance does not need it.

Compatibility & reproducibility

Measure semantics match jiwer, validated bit-identical on a 600-corpus golden set pinned in tests/.

agwer gratefully builds on and learns from these projects:

• RapidFuzz, C++ / FuzzyMatch swift edit distance

• jiwer defined the measure semantics and the easy, friendly API style this package follows.

• OpenAI Whisper created the English text normalizer that became the community standard for ASR evaluation. We vendor it faithfully under its MIT license, with attribution.

• NVIDIA NeMo text processing sets the bar for full text normalization across languages. Its semiotic class taxonomy guides our normalization roadmap.

License

Apache-2.0. Vendored Whisper normalizers: MIT (see src/agwer/normalizers/LICENSE_WHISPER).