Back

Benchmark report

Voice Code Bench: Exact Tokens in Speech-to-Text

A report on how speech-to-text systems handle exact structured values in English workplace speech, from emails and URLs to command-line flags, file paths, account identifiers, dates, and measurements.

Open Hugging FaceRead Paper

Problem

Structured tokens are easy to corrupt.

WER measures transcript similarity. Production voice interfaces need something stricter: the exact email, command, account identifier, URL, date, file path, dollar amount, product code, or measurement a downstream system must act on.

spokenDHCPtranscribed as DHP
spoken192.168.1.1one octet collapses
spokenP-I-N-Gnormalized to ping

WER

WER does not measure exact-token success.

Traditional WER treats all words as part of the same transcript surface. Production systems do not. A single wrong character in an email, CLI flag, file path, URL, account identifier, product code, or measurement can route the wrong ticket, fail an account lookup, or trigger the wrong command.

CTEMCanonical token/entity match

The fraction of target entities whose canonical value is recoverable from the ASR transcript.

TSRTask success rate

A recording succeeds only when every target entity in that recording is recovered correctly.

WERWord error rate

A supporting transcript-quality diagnostic, not the primary benchmark score.

Benchmark scope

What Voice Code Bench measures

Voice Code Bench is intentionally narrower than universal audio suites. It asks one product-critical question: did the transcript preserve every structured entity a downstream workflow needs?

Recordings300

human-recorded English WAV segments

Audio5.587h

34.9-122.9 sec clips

Target entities1,482

26 structured entity types

Domains8

workplace workflow domains

Results

Main findings

We evaluated 18 STT systems and rank them by CTEM. OpenAI GPT Live Transcribe leads CTEM. Deepgram Nova-3 and OpenAI GPT Live Transcribe tie for the highest TSR, while AssemblyAI Universal-3 Pro has the lowest WER. Best-in-column highlights show the strongest result for each metric. That is the benchmark's core point: entity-level recovery, complete task success, and general transcript quality are related but not interchangeable.

  1. 01

    OpenAI GPT Live Transcribe leads CTEM, showing that entity-level recovery and task-level success can disagree.

  2. 02

    Deepgram Nova-3 and OpenAI GPT Live Transcribe tie for the highest TSR; AssemblyAI Universal-3 Pro has the lowest WER.

  3. 03

    Commands, file paths, URLs, and environment variables expose large gaps across systems.

#1OpenAI GPT Live Transcribe (Streaming)
91.8%
#2ElevenLabs Scribe v2
91.6%
#3Deepgram Nova-3
90.9%
#4Cartesia Ink 2 (Streaming)
90.1%
#5OpenAI GPT Transcribe
89.3%
#6Deepgram Nova-3 (Streaming)
88.9%
#7Google Cloud Chirp 3
88.8%
#8Whisper Large-v3
87.6%
#8Meta Muse Voice Transcribe 1.0
87.6%
#10Meta Muse Voice Transcribe 1.0 (Streaming)
86.8%
#11Google Cloud Chirp 3 (Streaming)
86.1%
#12AssemblyAI Universal-3 Pro
84.5%
#13Thinking Machines Lab Inkling-NVFP4
84.3%
#14ElevenLabs Scribe v2 Realtime (Streaming)
84.0%
#15Inworld STT 1
83.7%
#16NVIDIA Parakeet TDT 0.6B v3
78.6%
#17AssemblyAI Universal-3 Pro Realtime (Streaming)
78.3%
#18Amazon Transcribe (Streaming)
75.2%
RankModelCTEMTSRWER
#1OpenAI GPT Live Transcribe (Streaming)Highest CTEM and joint-highest TSR in the released baseline suite, with 91.8% CTEM, 68.7% TSR, and 5.9% WER.91.8%68.7%5.9%
#2ElevenLabs Scribe v2Strong batch baseline with 91.6% CTEM, 67.7% TSR, and broad entity recovery.91.6%67.7%6.9%
#3Deepgram Nova-3Strong CTEM and joint-highest TSR in the released baseline suite, with 90.9% CTEM, 68.7% TSR, and 8.2% WER.90.9%68.7%8.2%
#4Cartesia Ink 2 (Streaming)Streaming-only baseline with 90.1% CTEM, 63.7% TSR, and 9.3% WER.90.1%63.7%9.3%
#5OpenAI GPT TranscribeBatch OpenAI baseline with 89.3% CTEM, 61.0% TSR, and 6.1% WER.89.3%61.0%6.1%
#6Deepgram Nova-3 (Streaming)Strong streaming baseline with 88.9% CTEM, 61.7% TSR, and 9.3% WER.88.9%61.7%9.3%
#7Google Cloud Chirp 3Strong email recovery and competitive TSR, but WER trails the top systems in this release.88.8%60.3%5.8%
#8Whisper Large-v3Open-source baseline with competitive TSR but weaker URL and command recovery.87.6%54.3%5.8%
#8Meta Muse Voice Transcribe 1.0Batch Meta baseline with 87.6% CTEM, 55.3% TSR, and 7.5% WER.87.6%55.3%7.5%
#10Meta Muse Voice Transcribe 1.0 (Streaming)Streaming Meta baseline with 86.8% CTEM, 55.7% TSR, and 7.4% WER.86.8%55.7%7.4%
#11Google Cloud Chirp 3 (Streaming)Streaming Google baseline with strong phone and IP recovery but lower overall TSR than batch Chirp 3.86.1%50.3%6.7%
#12AssemblyAI Universal-3 ProBatch AssemblyAI baseline with strong phone-extension and port-number recovery, but lower TSR than the top hosted systems.84.5%50.3%4.3%
#13Thinking Machines Lab Inkling-NVFP4Open-source batch baseline with 84.3% CTEM, 49.7% TSR, and 4.9% WER.84.3%49.7%4.9%
#14ElevenLabs Scribe v2 Realtime (Streaming)Realtime ElevenLabs baseline with strong person/team, version, and domain-term recovery but lower TSR than batch Scribe v2.84.0%46.3%6.8%
#15Inworld STT 1Batch Inworld baseline with 83.7% CTEM, 44.0% TSR, and 9.7% WER.83.7%44.0%9.7%
#16NVIDIA Parakeet TDT 0.6B v3Modal-hosted Parakeet baseline with 78.6% CTEM, 37.3% TSR, and 9.3% WER.78.6%37.3%9.3%
#17AssemblyAI Universal-3 Pro Realtime (Streaming)Streaming AssemblyAI baseline with 78.3% CTEM, 33.0% TSR, and 6.1% WER.78.3%33.0%6.1%
#18Amazon Transcribe (Streaming)Streaming Amazon baseline with 75.2% CTEM, 33.7% TSR, and 10.3% WER.75.2%33.7%10.3%

Entity heatmap

Where systems separate by entity type

Each cell is entity-level CTEM: exact recovery for that structured-value family. The color is relative within each column, making it possible to compare the systems on the tokens a particular product depends on.

Entity-level CTEMRows are STT systems. Columns are the structured token families in the benchmark.
Column leaderMiddleLargest gap
ModelEmailPhoneExtPerson /TeamAddressURLIP AddressPortCommandCLI FlagFile PathEnv VarCode SymbolVersionReference IDProduct CodeAccount /RecordCurrenc -yPctMeasure -mentNumberDateTimeAcronymSpelled SeqDomain Term
OpenAI GPT Live Transcribe (Streaming)75%98%100%94%48%66%96%100%78%98%71%91%91%93%95%93%92%100%100%100%100%100%98%98%96%95%
ElevenLabs Scribe v272%98%100%100%78%60%96%100%60%93%65%94%97%93%95%93%97%100%100%100%100%99%95%96%98%100%
Deepgram Nova-371%98%100%100%68%66%96%97%58%89%61%86%86%93%95%94%97%100%100%100%100%100%97%100%98%90%
Cartesia Ink 2 (Streaming)80%98%100%94%60%56%96%97%60%89%55%91%89%97%94%90%95%96%100%100%98%100%97%99%99%95%
OpenAI GPT Transcribe69%97%100%94%55%63%72%93%76%93%67%89%89%93%89%88%91%97%100%100%98%99%98%96%97%100%
Deepgram Nova-3 (Streaming)69%98%100%100%65%58%96%90%62%86%53%80%80%93%95%92%92%97%100%100%98%98%98%98%96%90%
Google Cloud Chirp 374%98%100%100%48%65%96%97%40%91%59%89%91%87%93%92%95%99%98%100%100%100%97%95%92%95%
Whisper Large-v374%97%100%94%60%42%92%93%46%86%55%89%91%93%95%89%91%97%100%100%97%97%93%98%97%85%
Meta Muse Voice Transcribe 1.069%98%100%94%55%40%92%97%52%95%53%91%94%87%93%88%91%88%98%100%100%100%100%98%96%90%
Meta Muse Voice Transcribe 1.0 (Streaming)60%98%100%100%60%34%92%100%54%91%51%91%94%93%93%83%91%91%100%100%100%99%98%96%95%85%
Google Cloud Chirp 3 (Streaming)63%98%100%100%48%58%96%97%50%86%55%86%83%87%91%93%88%88%98%95%97%99%92%94%93%90%
AssemblyAI Universal-3 Pro51%88%97%94%68%47%80%97%44%80%63%51%74%87%91%84%91%92%98%100%98%100%98%98%95%100%
Thinking Machines Lab Inkling-NVFP455%92%100%83%40%48%76%83%58%75%61%77%77%93%91%87%86%96%98%100%98%100%97%94%89%95%
ElevenLabs Scribe v2 Realtime (Streaming)63%88%93%100%57%34%92%97%44%82%45%54%91%100%91%80%92%89%100%98%95%97%93%98%97%100%
Inworld STT 158%92%97%94%50%34%68%83%54%84%51%91%83%97%86%90%88%92%100%95%100%96%97%96%91%85%
NVIDIA Parakeet TDT 0.6B v332%83%100%94%45%40%76%83%32%70%33%86%86%97%78%77%78%88%100%100%98%97%92%98%90%90%
AssemblyAI Universal-3 Pro Realtime (Streaming)46%82%87%94%55%34%68%90%36%89%47%66%77%93%76%79%77%79%100%97%95%98%98%93%85%100%
Amazon Transcribe (Streaming)48%98%100%83%45%10%88%93%8%45%8%43%54%80%84%84%86%75%100%95%97%99%98%94%95%80%

CTEM vs WER

Entity recovery and word error rate are not interchangeable

WER stays visible as a transcript-quality diagnostic. CTEM directly measures how often the canonical entities required by downstream workflows survive.

Highest CTEMOpenAI GPT Live Transcribe (Streaming)

91.8% canonical entity match

Lowest WERAssemblyAI Universal-3 Pro

0.0432 word error rate

Interpretation7.2 pp gap

The lowest-WER model trails the top entity-recovery model by this many percentage points on canonical entity match.

Metrics

How to read CTEM, TSR, and WER

CTEM shows the global exact-match rate across entities. TSR is stricter at the recording level because one missed entity can invalidate the task, while WER remains a supporting diagnostic for transcript quality.

  • Low WER does not guarantee usable transcripts when the application depends on exact structured values.
  • OpenAI GPT Live Transcribe has the highest CTEM; it and Deepgram Nova-3 tie for the highest TSR.
  • Thinking Machines Inkling reaches 84.3% CTEM and 49.7% TSR as an open-source batch baseline.
  • Commands, file paths, URLs, and environment variables expose large gaps across systems.

Dataset composition

Recording and entity distribution

Voice Code Bench is a test-only evaluation dataset, not a training corpus. The release covers 49 scenarios across four difficulty levels.

Entity Coverage

Reference ID
150
Spelled sequence
97
Product code
90
Date
89
Acronym/initialism
85
Currency amount
75
Email address
65
Plain number
65
Account/record #
65
URL
62
Measurement
61
Phone number
60
Time
60
File path
51
Command
50
Percentage
50
CLI flag
44
Postal address
40
Environment variable
35
Code symbol
35
Phone extension
30
Version
30
Port number
30
IP address
25
Domain term
20
Person/team name
18

Domain Coverage

Technical IT/dev
55
Contact routing
45
Retail/logistics/order
45
Finance/billing
40
Healthcare admin
35
Legal/insurance/gov
35
Education/workplace
25
Dense mixed stress
20

Duration Buckets

Under 40 sec
5 (2%)
40-60 sec
102 (34%)
60-90 sec
165 (55%)
90-120 sec
27 (9%)
120+ sec
1 (0%)

Recording Length Summary

Mean length67.0 sec
Median length65.5 sec
Range34.9-122.9 sec
Mean entities4.9 / recording
Entity range3-8 / recording

Speaker metadata

Who is represented in the benchmark

Speaker ID, sex, accent, and age-bucket metadata are available for all 300 recordings. The current release includes 85 anonymized speaker IDs.

FieldCategoryRecordingsShare
Speaker sexFemale152
Male144
Non-binary4
Age bucket36-45102
18-2576
46-5554
26-3553
56-6510
66+5
AccentAmerican general183
American New York18
Nigerian16
American California14
American Midwest14
American Southern14
British RP13
Indian6
Neutral6
Canadian5
Spanish Mexican5
British Northern2
Kenyan2
Eastern European1
Russian Moscow1

Audio quality

Audio-quality measurements for every recording

Signal-to-noise ratio, background-noise level, speech level, loudness, and click/pop metadata are available for all 300 recordings.

Quality Field Summary

Mean and median values across all recordings.

MetricMeanMedianRange
Signal-to-noise ratio61.8 dB61.1 dB36.5 to 108.8 dB
Background noise RMS-80.6 dBFS-80.1 dBFS-120.0 to -54.9 dBFS
Speech RMS-18.8 dBFS-18.8 dBFS-27.4 to -10.1 dBFS
Integrated loudness-23.0 LUFS-23.0 LUFS-32.0 to -15.0 LUFS
Click/pop events0.11 / min0.00 / min0.00 to 1.98 / min

Metadata Summary

Audio quality metadata300 / 300
Mean SNR61.8 dB
Median SNR61.1 dB
SNR range36.5-108.8 dB
Mean loudness-23.0 LUFS
Median loudness-23.0 LUFS

SNR Buckets

80+ dB
22 (7%)
60-80 dB
142 (47%)
45-60 dB
118 (39%)
Under 45 dB
18 (6%)

Scoring method

From raw audio to an auditable benchmark score

1

Audio to transcript

The provider receives no gold transcript or benchmark entity values.

2

Transcript to entities

The released scorer uses an LLM-assisted recoverability verifier and stores evidence and reasons.

3

Entities to benchmark

CTEM, TSR, entity CTEM, WER, score files, and verifier-provenance outputs are generated.

Three transcript layers

Each item includes template, acoustic, and canonical transcript layers.

Raw audio only

Systems receive only the audio file in the main setting, without benchmark prompts, entity lists, domain labels, custom vocabulary, grammar constraints, or post-ASR correction.

Canonical recovery

Scoring asks whether the written value a downstream application needs is recoverable from the ASR transcript.

Verifier audit trail

Entity scoring uses an LLM-assisted recoverability verifier with versioned prompts, response schemas, evidence, reasons, and a released audit sample.

Diagnostic use

VoiceCodeBench is intended for diagnostic ASR evaluation, provider comparison, regression tracking, and per-entity risk analysis.

Out-of-scope uses

The release is not intended for training, fine-tuning, speaker identification, biometric modeling, voice cloning, or demographic profiling.

Reproducibility

Files, baseline IDs, and commands

The release includes the artifacts needed to inspect the evaluation and reproduce its checked-in results.

Repository layout

  • data/audio/001.wav through data/audio/300.wav
  • data/metadata.json
  • baselines/predictions/*.json
  • baselines/results.csv
  • audit/verifier_audit_samples.csv
  • scripts/voice_code_bench/
  • paper/voice-code-bench.pdf
  • DATASET_CARD.md

Baseline IDs

  • openai_gpt_live_transcribe_streaming
  • elevenlabs_scribe_v2
  • deepgram_nova3
  • cartesia_ink_2_streaming
  • openai_gpt_transcribe
  • deepgram_nova3_streaming
  • google_cloud_chirp_3
  • whisper_large_v3
  • meta_muse_voice_transcribe_1_0
  • meta_muse_voice_transcribe_1_0_streaming
  • google_cloud_chirp_3_streaming
  • assemblyai_universal_3_pro
  • modal_inkling
  • elevenlabs_scribe_v2_realtime_streaming
  • inworld_stt_1
  • modal_nvidia_parakeet_tdt_0_6b_v3
  • assemblyai_universal_3_pro_streaming
  • amazon_transcribe_streaming

Commands

Install scorer
python -m pip install -e .
Reproduce release artifacts
./scripts/reproduce_release.sh
Run tests
python -m pytest
Run new provider baselines
vcb-run --stt-mode all --output-dir runs/full-local

Paper

Technical paper for VoiceCodeBench, including the benchmark motivation, scoring protocol, baseline setup, and release details.

Read paper PDF

Context

Why this benchmark exists

Broad voice benchmarks measure general capability. Voice Code Bench is narrower on purpose: it isolates exact structured-token recovery, the failure mode that matters when speech becomes software input.

SUPERB / ML-SUPERB

Name the task, source dataset, split, and metric explicitly.

Voice Code Bench exposes its 26 entity types and scoring rules up front, so the evaluation target is clear before the results table.
Dynamic-SUPERB Phase Two

Breadth creates authority but makes one-score summaries fragile.

Voice Code Bench is intentionally narrow: it measures exact structured-token recovery rather than claiming broad audio-language coverage.
AIR-Bench / AudioBench

Instruction-following audio systems require hybrid scoring.

The LLM verifier is limited to entity-presence judgments, while canonical targets, evidence, and review files remain auditable.
AHELM

Fairness, toxicity, safety, and robustness belong inside the benchmark frame.

Voice Code Bench surfaces speaker metadata completeness and known imbalance as benchmark context, not appendix material.
Hugging Face dataset cards

Good documentation is decision support, not decoration.

The benchmark page puts scale, provenance, schema, metadata, limitations, and prohibited uses where practitioners can evaluate fit quickly.

Failure types

Common transcription failures

The purpose is not only to rank STT providers. It is to make the recurring failure patterns visible enough for product and research teams to fix.

Omissionthe entity disappears from the transcript
SubstitutionDHCP becomes DHP
Formatting driftan email, phone number, or currency amount is normalized incorrectly
Spelling collapseP-I-N-G becomes ping
Near misssemantically close but unusable exact value
Extra entitya structured token appears that was never spoken

Dataset documentation

Dataset documentation requirements

Good benchmark documentation is not decorative. It should make scale, schema, metadata, risks, and scoring rules obvious before someone downloads the data or cites the result.

Use cases

Where exact token recovery matters

Voice coding, support operations, CRM dictation, technical education, finance, healthcare, and legal workflows all depend on exact structured speech.

  • 01ASR vendor selection
  • 02Voice coding
  • 03Support QA
  • 04CRM and form dictation
  • 05Technical troubleshooting
  • 06Agentic voice workflows

Limitations

Current limitations

  • Current release is English-only and focused on compact workplace-style dictation.
  • It does not cover meetings, overlapping speech, casual conversation, broadcast audio, voice search, noisy field recordings, telephony codecs, or long-form dictation.
  • Synthetic content may not capture every distributional property of real production workflows.
  • The LLM-assisted verifier introduces dependence on a verifier model and prompt.
  • Commercial ASR rankings are tied to the evaluation dates and provider behavior recorded in the baseline artifacts.

Roadmap

Planned extensions

The benchmark should grow in entity types, model coverage, speaker/environment diversity, and auditable slice reporting while preserving the exact-token focus.

  1. 01

    Versioned releases that preserve prior dataset revisions

  2. 02

    Multilingual structured-token recovery

  3. 03

    Noisier environments and broader accent coverage

  4. 04

    Broader conversation formats beyond compact dictation

  5. 05

    Documented changes to entity taxonomy, scoring policy, verifier versions, and baseline systems

  6. 06

    Human verifier calibration and expanded audit samples

Next

Dataset and paper

The dataset, baseline outputs, evaluation harness, and documentation are available on Hugging Face. The paper records the motivation, benchmark design, scoring protocol, baseline setup, and release results.

Open Hugging FaceRead Paper