Converse-STT Technical Report

The benchmark

Converse-STT measures how accurately speech-to-text systems transcribe unscripted, two-person American English conversation when each speaker is recorded on their own microphone for a full 15-minute session. This report documents the benchmark end to end: how the recordings were collected and transcribed, how the audio reaches each model, how text is normalized, how word and character error rates are computed and pooled, what the 19 systems on the leaderboard scored, and what the errors actually are.

Public ASR leaderboards are dominated by short, single-speaker clips. Conversation is different: speakers overlap, backchannel (“yeah”, “right”, “mm-hmm”), restart words mid-syllable, and talk for a quarter of an hour without a break. Converse-STT Version 1, scored in September 2026 with Cekura^[9], showed that twelve of fifteen models had a higher error rate on Ocular AI’s conversational recordings than on public short clips, and that the leading model changed with the dataset.

Version 2 keeps the same four recordings and makes the evaluation reproducible by anyone. Every number on the board is produced by the open-source harness (MIT): the dataset manifest, the per-model inference configuration, the normalizer, the scorer and the report are all code, and every run directory keeps the raw transcripts and the normalized texts that produced its score. The design follows established conventions for conversational ASR evaluation: a separate channel per speaker, verbatim references^[2], Whisper English normalization^[1], and summed edit counts, so that a model’s number here can be read next to numbers from other public harnesses that use the same conventions. The implementation is independent and deliberately small: one Python package, one manifest format, one command per step.

Scoring convention

A word error rate is only comparable with another when the normalizer, the reference convention, the pooling rule and the audio each system received are the same. Every number on the leaderboard and in the findings uses the convention below. The sections after it are how the reference was built and how the percentage was computed.

Dataset collection

The audio is the public four-conversation sample of Ocular AI’s American English Hi-Fi full-duplex collection, published on Hugging Face^[3] under CC BY 4.0. The harness pins revision 29b74ad4.

Setting and prompting

Two participants hold a conversation on Ocular’s recording platform. Each participant is captured on a separate microphone into a separate file for the entire session, so there is no mixing and no voice-activity gating between speakers at the recording stage. Each session opens with a supplied discussion prompt that one participant reads or paraphrases (for example, “Consider the ethical and practical implications of using drones in everyday life…”). The prompt is part of the recording and of the reference; everything after it is unscripted.

Participants and sessions

Eight distinct speakers, all self-reporting American English as accent and English as native language; five female, three male; self-reported age bands 35–45 (one), 45–55 (three) and not reported (four). Demographics are self-reported and were not verified. The sessions were recorded on 12 and 15 August 2026 (UTC). Topics: drones and privacy; interpreting news sources; brand loyalty and customer engagement; fitness plans.

Synchronization

The two files of a session do not start at the same instant. Each filename carries the capture start in epoch milliseconds (<room_start_ms>-<speaker_id>-hifi-audio-<capture_start_ms>.flac); the later track in each session starts 1.298–2.125 s after the earlier one. The harness derives this offset from the filenames and shifts the later track’s word timestamps onto a shared conversation timeline.

Transcription protocol

According to the collection’s documentation, references were produced by an ASR-assisted first pass followed by human transcription correction, a second human review, and an Ocular audit. Word-level timestamps were produced afterwards by forced alignment of the reviewed text to the audio.

Audio specification

All eight files are mono 24-bit PCM FLAC at 48 kHz, delivered as recorded, with no resampling, trimming or loudness processing. The harness never alters the archived audio; conversions for upload are made from it per request and discarded afterwards.

Transcripts and reference conventions

References are verbatim. They keep fillers, repetitions, false starts and word fragments, and annotate non-speech events with SGML-style tags. The references contain no digits at all; every number is spelled out as spoken. Contractions are written as spoken. The dataset card notes known imperfections: two segments without word arrays, one segment ending 69 ms after its track, and alignment that is less reliable for short fillers. None of these affect text scoring, which uses the full-track transcript, not the word arrays.

Dataset statistics

Speech time below is computed from the word-level timestamps after merging same-speaker words separated by at most 200 ms, the pause-merging threshold used in multi-speaker ASR corpora such as CHiME-6^[2]. Overlap is the time both speakers are active, as a share of total speech time. Turns are same-speaker stretches separated by more than one second of that speaker’s silence.

From the published files to a manifest

ocular-evals dataset pull downloads train/metadata.jsonl and the eight FLACs from the pinned revision and writes data/converse-en-sample/manifest.json, the only dataset artefact the rest of the harness reads. For each session it groups the two rows by session_id; parses the capture start from each filename, orders tracks by capture time (then speaker_id), names them ch1 and ch2, and records offset_s (0.000 / 2.125, 0.000 / 1.321, 0.000 / 1.298, 0.000 / 2.109 s); copies each FLAC byte-for-byte and records its SHA-256, duration and sample rate; stores the full transcript string as the reference, and the word entries with start and end shifted by offset_s; and keeps the self-reported metadata fields.

The manifest records the Hugging Face repo, revision and the SHA-256 of metadata.jsonl. The dataset’s identity for the rest of the pipeline is a 16-hex-character digest of the sorted (conversation, track, sha256) triples, cea63205dcfdfced for this sample; a run refuses to resume or re-score if that digest changes, and every track’s checksum is re-verified immediately before it is uploaded. ocular-evals dataset verify re-checks all files against the manifest.

Task definition

The task is batch, single-channel, single-speaker transcription of a 15-minute track with a language hint of en. Each of the eight tracks is sent to the system whole, or in the largest pieces the system’s API allows, and the returned text is scored against that track’s reference. There is no mixing of the two speakers, no diarization requirement, and no streaming; the score is about what the system writes down for one person talking in a real conversation, with the other person’s voice occasionally audible through the microphone.

Scoring per separate channel is the standard “separated channels” condition for dual-microphone conversational corpora^[2]. A mixed-channel, speaker-attributed condition (cpWER) is on the harness roadmap and is not part of this report; who spoke when is the sibling benchmark, Converse-Diarization, on the same recordings.

Inference harness

ocular-evals run --model <name> resolves the entry in models.toml to a provider adapter and processes the selected tracks sequentially. Eight adapters cover the systems in this report.

Text normalization

Both the reference and the hypothesis pass through the same function before any comparison. The profile used for every number in this report is whisper: six steps, the fourth of which is the English text normalizer published with Whisper^[1] (whisper-normalizer 0.1.12).

Scoring

For each track the normalized reference and hypothesis are tokenized two ways: words are whitespace-separated tokens; characters are every non-whitespace character. Each pair of token sequences is aligned with jiwer 4.0^[6] using identity transforms, which is a standard Levenshtein alignment^[4] with unit cost for substitution, deletion and insertion. The result is stored as hits H, substitutions S, deletions D and insertions I. Rates are derived, never stored.

Pooling by counts means a short track cannot carry the same weight as a long one, and a catastrophic track contributes exactly its error count. WER is unbounded above^[5]; the pre-fix Voxtral run scored 107% because it produced 12,800 inserted words on one track. A track with no hypothesis (failed after retries) is excluded from the sums and the run is marked incomplete; every run on the board is complete, 8 of 8 tracks. A hypothesis that is empty text scores as all deletions.

CER is reported alongside WER as a tiebreaker and as a check that WER is not being driven by tokenization (for example compounding “pushups”); the primary metric for English is WER. Wall-clock latency per track is recorded and summarized as the median across the eight tracks. It is an end-to-end batch number including upload and any proxy hop and is not comparable across routes; it is reported for context, not ranked.

Reproducibility and audit trail

Each run directory holds run.json (harness version, full resolved model config, model digest, dataset digest and source, normalization profile, timestamps), hypotheses.jsonl (one row per transcribed track with raw text, segments, latency, audio SHA-256 and the full raw API response), failures.jsonl, scores.jsonl (per-track WER/CER counts and the exact normalized reference and hypothesis strings that were aligned) and summary.json (pooled counts and rates, per-conversation breakdown). The leaderboard on this site is rendered from those summaries, and the transcript panel behind each system is the scores.jsonl alignment.

Streaming systems

Converse-STT scores a finished transcript of a finished recording; it is a batch benchmark and does not measure latency, partial-result stability or endpointing. Several systems on the board also exist as streaming products (Deepgram, AssemblyAI, ElevenLabs, xAI and Reson8 all sell real-time endpoints), and for those the board uses the vendor’s batch endpoint, as a vendor evaluation would. Two systems, NVIDIA’s Nemotron 3 ASR and Nemotron 3.5 ASR, exist only as streaming models: cache-aware FastConformer-RNNT encoders^[7] in which the right attention context (lookahead, in 80 ms frames) fixes how much future audio each frame may see. Because the streaming steps never overlap and the cache reproduces the full left context, offline inference at a given lookahead yields the same transcript the streaming path produces at that chunk size; the lookahead is therefore the single knob that moves these models along their latency–accuracy curve (80 ms, 160 ms, 320 ms, 560 ms, 1.12 s).

The policy is: a streaming-only model’s board row is its highest-accuracy operating point (lookahead 13 frames = 1.12 s), which is the configuration a batch deployment would use and the one comparable to the other systems’ batch endpoints. Lower-latency operating points are not ranked; they are reported in a separate figure so the accuracy cost of latency is visible without putting an 80 ms streaming configuration in the same ordering as a batch API. One streaming system is scored through its streaming path on purpose: Version 1’s “GPT Realtime Whisper” is the OpenAI Realtime API’s transcription session (server VAD, whisper-1), a product that exists only as a stream, and reproducing the Version 1 row means running it that way. Its score should be read as the accuracy of that product on this audio, not of whisper-1 as a batch decoder.

Evaluated systems

The leaderboard ranks two sets of systems. First, each vendor’s latest general-purpose transcription model: where a vendor offers several current models we take the newest generation (GPT Transcribe; Gemini 3.5 Transcribe; Grok Voice Transcribe 2.0 rather than Grok STT 1.0; MAI-Transcribe 2 rather than 1.5), the flagship size within a generation (Qwen3-ASR 1.7B rather than 0.6B), and the general-domain model rather than a domain variant (Omi General rather than Omi Medical). Second, every model that was scored in Converse-STT Version 1 and that the credentials we hold can reach, so that the Version 2 board can be read against Version 1 model for model: this adds Chirp 3, GPT-4o Transcribe, GPT-4o mini Transcribe, GPT Realtime Whisper and Reson8 to the vendor-latest set (Nova-3 and Universal-3.5 Pro are in both sets; Reson8 is both the Version 1 leader and its vendor’s only model). As a result Google has two rows and OpenAI four; within a vendor the rows are independent runs and are ranked independently.

NVIDIA has three rows for a different reason: Parakeet TDT 0.6B v3^[8] is its hosted batch model, while Nemotron 3 ASR (English) and Nemotron 3.5 ASR (multilingual) are a separate, streaming-only family that NVIDIA positions as a complementary product rather than a successor; both run locally at their batch-quality setting. Sibling and older-generation models on neither list (Grok STT 1.0, MAI-Transcribe 1.5, Qwen3-ASR 0.6B, Omi Medical, GPT-4o Transcribe Diarize, Whisper large-v3) were run with the same harness on the same data and are referred to below where they illuminate a finding, but they are not ranked.

19 systems are on the board, all scored against the same pinned manifest. Nine were called on the vendor’s own endpoint with the vendor’s key; two were run locally with open weights; eight through OpenRouter’s /audio/transcriptions gateway, which exposes a catalogue of transcription models behind one OpenAI-style API and one key. OpenRouter adds a proxy hop to latency, may route an id to more than one upstream host, and does not forward every vendor-specific parameter (notably timestamp_granularities); the board says “via OpenRouter” on those rows.

Leaderboard

Pooled over 12,555 reference words and 47,784 reference characters. Scribe v2 leads at 1.22%; GPT-4o Transcribe is last at 13.82%. One WER point is 125.6 words and 0.1 point is about 13 words, so differences under roughly 0.3 points between adjacent systems should be read as ties. Pairs that are effectively tied on this data: Reson8 and Grok Voice Transcribe 2.0 (2 words apart), Gemini 3.5 Transcribe and Muse (23 words), Nova-3 and Omi General (5 words), Chirp 3 and MAI-Transcribe 2 (3 words), Parakeet and GPT Transcribe (5 words), Qwen3-ASR 1.7B, GPT-4o mini Transcribe and Universal-3.5 Pro (10 and 2 words), Voxtral and GPT Realtime Whisper (15 words).

Key findings

Scribe v2 is alone at the top. 1.22% is 1.7 points (214 words) ahead of second place, which is itself a tie: Reson8 (2.92%) and Grok Voice Transcribe 2.0 (2.94%) are two words apart. Its 153 total errors are spread thinly: the single largest category is 29 fragment deletions, and it transcribes the partner-bleed track with zero cross-talk insertions.

The spread is wider than on short-clip benchmarks. Nineteen models span 1.2–13.8%, a factor of 11 between best and worst (1.2–6.7%, a factor of 5.5, without GPT-4o Transcribe); public short-clip boards span under a factor of two. Long-form conversational audio exposes failure modes short clips never trigger: phrase-level omission (GPT Transcribe, GPT-4o Transcribe, Voxtral), decoder loops (Voxtral without segment decoding), hallucination on short VAD-cut turns (GPT Realtime Whisper), language switching on backchannels (Qwen3-ASR’s 嗯, the GPT-4o family’s ja), and cross-talk transcription (Parakeet, MAI-Transcribe 2, AssemblyAI).

Deletions dominate for most systems; what gets deleted is conversational. For 16 of 19 systems D is the largest single category, and for 10 (Gemini, Muse, both Nemotron models, GPT Transcribe, GPT-4o Transcribe, GPT-4o mini, GPT Realtime Whisper, Universal-3.5 Pro, Voxtral) D exceeds S + I. The exceptions are Scribe v2, where substitutions lead at a very low level, and Parakeet and Qwen3-ASR, where insertions lead. The most-deleted words corpus-wide are i, a, and, it, you, know, yeah, right, like: fragments, repeated function words and backchannels. Several systems appear to apply disfluency cleanup by design: Universal-3.5 Pro deletes 105 “you”/“know” tokens, roughly 50 of the about 300 “you know” in the corpus; MAI-Transcribe 2 drops isolated “yeah”/“right” backchannels (48) and GPT-4o mini Transcribe does the same (42). Nemotron 3.5 ASR behaves the same way (74 deleted “yeah”/“right”), while its English sibling deletes 27.

The hard track is the quiet one. 0fcd38f2/ch1, the male speaker whose channel is gated and whose partner talks 70% more than he does, is the worst or second-worst track for 10 of 19 systems, and conversation 0fcd38f2 is the worst conversation for 12. Insertions of the partner’s speech account for 112 of Parakeet’s 130 insertions there, and 44–45 for MAI-Transcribe 2 and AssemblyAI. Systems that ignore faint background speech (Scribe, Reson8, Grok, Gemini, Chirp 3, Muse, Nova-3, Omi, GPT Transcribe, GPT-4o Transcribe, GPT Realtime Whisper, Qwen, Voxtral, both Nemotron models) insert nothing from it; Nemotron 3 ASR’s 2.78% on that track is its second-best.

Newer is not always better within a vendor. OpenAI’s rows are the clearest case: GPT Transcribe 4.58%, GPT-4o mini Transcribe 5.11%, GPT Realtime Whisper 6.71% and GPT-4o Transcribe 13.82%, the last 9.2 points behind the current model on the same audio (the unranked GPT-4o Transcribe Diarize scored 5.53% and open-weight Whisper large-v3 6.32%). Google’s two rows are closer: Chirp 3 (3.72%) is within 0.6 points of Gemini 3.5 Transcribe (3.13%). The retired siblings are mostly worse than their successors (MAI-Transcribe 1.5 3.97% against 3.74%; Qwen3-ASR 0.6B 5.41% against 5.03%), with one exception: Grok STT 1.0 (2.82%) is 13 words better than Grok Voice Transcribe 2.0 (2.94%), a tie at this resolution.

Within NVIDIA’s open models, the English streaming model beats the hosted batch model. Nemotron 3 ASR (EN) at its 1.12 s setting scores 4.11% against Parakeet TDT 0.6B v3’s 4.54% with a quarter of the insertions (52 against 222: no cross-talk, no mhmm-style fillers), and the multilingual Nemotron 3.5 ASR trails both at 5.87%, mostly through deletions (488), which matches NVIDIA’s own guidance to prefer the English model for English.

Configuration decides whether a score is about the model. Voxtral scored 107.7% with a plain json request and 6.59% with timestamp_granularities=segment. With OpenAI’s default request, GPT-4o Transcribe scored 21.87% and GPT-4o mini 21.62% on this data; with chunking_strategy=auto, the board’s 13.82% and 5.11%.

Per-track word error rate

86c27418/ch1 (a female speaker, 45–55, on the fitness topic) is the easiest track for 14 of 19 systems and below 3.6% for every system except GPT Realtime Whisper (3.81%) and GPT-4o Transcribe (7.73%). Track difficulty is driven more by the speaker and channel than by the topic: conversation 0fcd38f2 is the hardest for 12 systems although drones and deliveries are the least technical subject in the set. Ten track scores are above 10%. Five belong to GPT-4o Transcribe, one is GPT Realtime Whisper on 2bb0cfe2/ch1 (11.04%), and the other four (Parakeet and Universal-3.5 Pro on 0fcd38f2/ch1, Universal-3.5 Pro on 86c27418/ch2, Voxtral on ea217639/ch2) each have a single identifiable cause: cross-talk, discourse-marker deletion, and a residual repetition loop.

Anatomy of the errors

The alignment in scores.jsonl lets every error be attributed. Grouping the 19 runs’ substitutions, deletions and insertions by cause:

Verbatim-reference costs shared by every system

Fragment restarts (“I-”, “s-”, “th-”) are almost never written by any model; deletions of one- and two-letter fragment tokens range from 29 (Scribe) through 40–140 (most systems) to 162 (GPT Transcribe) and 179 (Voxtral), with GPT-4o Transcribe far above at roughly 350 because it drops whole phrases rather than fragments alone. The hyphenated-letter convention (“A-I”) costs 10–15 errors per model. These are properties of the benchmark, applied uniformly; they set a floor of roughly 0.3–0.5% below which no system that writes standard orthography can go.

Backchannels and discourse markers removed by the system

Aligned examples

These are passages from the normalized reference with one system’s edits marked from the alignment the score was computed on: a substituted word shows the model’s word with the reference word struck through, a deleted word is struck, an inserted word is highlighted. Choose a passage, then switch systems to compare how each handles the same seconds of speech.

Case studies in configuration

Voxtral Mini Transcribe 2 and the repetition loop

With the plain json request Voxtral transcribes ea217639/ch2 correctly for about three minutes, then emits a 75-word passage (“…and I think that’s a great way to incentivize a lot of people…”) 179 times, recovers, loops once more, and finishes correctly: 14,262 hypothesis words against a 1,544-word reference, 12,800 insertions, track WER 835%, corpus WER 107.74%. The loop is deterministic across retries, with and without language, and at temperature=0; it does not occur on the original 48 kHz FLAC upload or when the request asks for segment timestamps (timestamp_granularities=segment), which makes the server decode segment by segment. The board uses segment decoding: 6.59%. A residual remains on the same track (1,599 words, 138 insertions, the passage repeated three times, 14.90%); it is reported as scored. The OpenRouter route for the same model cannot carry the parameter and loops identically (107.72%), so it is not on the board.

OpenAI server-side chunking

The GPT-4o-family models returned transcripts that stop or skip for long stretches on 15-minute files unless the request sets chunking_strategy=auto; the board rows use it.

Relation to Converse-STT 1.0

Converse-STT Version 1 scored 15 systems on the same four recordings with Cekura’s pipeline in September 2026^[9]. It stays published on its own page; Version 2 scores are not comparable with it and are not carried over. Seven of the 15 have been re-run with the Version 2 harness and are ranked on this board as ordinary runs.

Relation to other public boards

Seven of the vendor-latest board systems are also on a public short-clip US-English STT leaderboard (3,972 short clips, clean to noisy). Six of seven score lower here, by 0.04 to 3.6 points, consistent with close-talk audio; the rank correlation across the seven is nil (Spearman ρ = −0.14). Scribe v2 is 5th there and 1st here; Qwen3-ASR is 3rd there and 10th here; Grok Voice Transcribe 2.0 is 10th there and 2nd here. The disagreement is expected: a short-clip board measures per-utterance recognition under varied acoustics, while Converse-STT measures sustained transcription of one speaker in a live exchange, where omission, looping and cross-talk handling dominate the error budget.

Streaming operating points

The two Nemotron models were re-run at lower lookaheads with nothing else changed: same weights, chunking, prompt and hardware. Each row is a full pass over the eight tracks, scored exactly like a board run, but only the 1.12 s rows are ranked.

Limitations

The leaderboard is a ranking on this sample, under the convention above. It is not a result for every room, language or microphone a transcription system will meet.

Size and data

Four conversations, eight speakers, 12,555 words. A 0.1-point difference is 13 words; adjacent ranks are often not distinguishable. The board is a reproducible measurement on a fixed public sample, not a population estimate. The private Ocular collection the sample is drawn from is larger; this report covers only the public sample. The recordings have been public since the Version 1 release and cannot be assumed unseen by any system. References were human-corrected and audited but are not claimed error-free; a handful of residual reference errors would affect every system equally.

Routes and runtimes

Eight of the nineteen board systems were reached through OpenRouter. The proxy may select among upstream hosts, adds latency, and does not forward every parameter. Five systems were chunked client-side because of request or model limits on their route, and GPT Realtime Whisper is segmented by the server’s VAD; a vendor’s batch endpoint might behave differently. The two Nemotron rows were produced on our hardware with the Transformers port of NVIDIA’s checkpoints rather than NVIDIA’s NeMo/Riva runtime or hosted NIM. The weights are the same; decoder implementation details may differ slightly from NVIDIA’s reference numbers, and their latency column is compute time, not an API round trip.

Board composition and configuration

The board mixes each vendor’s current model with older models kept for continuity with Version 1, so a vendor’s position depends on which row is read: OpenAI spans 4.58% to 13.82%. Rows are independent runs; no vendor aggregate is computed. A single request parameter can move a score by 8–100 points; we used documented defaults plus the minimum settings needed to process long audio, and recorded every parameter; other reasonable choices exist and were not explored.

Normalizer and scope

Whisper’s English normalizer removes a fixed filler list and does not neutralize variant spellings (mhmm, huh, 嗯), splits hyphenated letter sequences, and maps “oh” to 0. These are symmetric but visible: about 0.5 points for Nova-3, about 1.7 for Qwen3-ASR 1.7B. The benchmark is English only, batch scoring only, separated channels only: no diarization, no mixed-channel cpWER, no latency ranking, no cost. These are roadmap items for the harness.

Reproducing a score

The harness is open source on GitHub. It downloads the pinned revision, verifies the FLACs, runs a model and renders the report.

uv sync --group dev
cp .env.example .env                      # add the keys you hold; names only ever appear in models.toml
uv run ocular-evals dataset pull          # Hugging Face sample @ pinned revision -> data/converse-en-sample
uv run ocular-evals dataset verify        # SHA-256 + duration of every track
uv run ocular-evals check                 # credentials and dependencies per model
uv run ocular-evals run --model elevenlabs-scribe-v2
uv run ocular-evals report --board converse-stt   # -> reports/converse-stt.html

A new OpenAI-protocol model is one models.toml entry. Models behind other protocols need an adapter implementing Provider.transcribe(path, language) → Hypothesis; the xai, meta and reson8 adapters are 170–200 lines each and are the templates for hosted APIs, nemotron-asr for local open-weight models. Changing anything that affects output changes the model digest and requires --fresh; changing only transport settings does not. The test suite (111 tests) runs without network access and covers the manifest builder, normalizer, metrics, chunker, retry logic and every adapter against mocked HTTP responses.

Conclusion

On four unscripted American English conversations, pooled word error rate after Whisper English normalization puts ElevenLabs Scribe v2 alone at the top at 1.22%, 214 words ahead of a tie between Reson8 (2.92%) and Grok Voice Transcribe 2.0 (2.94%). Nineteen systems span a factor of eleven, and the systems at the bottom are not there because they misheard words: they are there because they left words out, looped, hallucinated on short turns, or transcribed the other speaker.

The errors that separate systems are conversational. Fragments, repeated function words, and backchannels are what gets deleted; variant filler spellings and the partner’s voice through a gated microphone are what gets inserted. A single request parameter moved two systems by more than the whole spread of the top ten, which is why every parameter is recorded and every alignment is published.

The number to quote from the leaderboard is the pooled WER under this convention: verbatim reference, Whisper normalization, one track per speaker, all eight tracks summed. Read the ranking as a result on this sample, under that convention, and open a system to read the transcript behind its score.

License and citation

The harness is MIT-licensed. The audio and references are Ocular AI’s and are released under CC BY 4.0 on Hugging Face; model outputs and scores you produce are yours. The sample is an evaluation set: do not train, fine-tune or distil on it. Contact the research team with questions.

@techreport{ocularai2026converse-stt,
  title       = {Converse-STT 2.0: Benchmarking Speech-to-Text on Real Two-Person American English Conversation},
  author      = {{Ocular AI}},
  institution = {Ocular AI},
  year        = {2026},
  month       = {10},
  url         = {https://www.useocular.com/converse/stt},
}

References