This test set contains real telephone conversations that were anonymized: personal data was silenced and every voice was replaced. By requesting access you agree to use it for evaluation and research only, not to train or fine-tune models on it, and not to attempt to identify any speaker, person or organization in it.

Armenian Telephone Overlap

A test set for overlapped, narrow-band (8 kHz) conversational Armenian. 29 clips · 11.1 minutes · 1574 reference words, 355 of them spoken while the other talker is also speaking · 0.7 minutes of two-talker overlap.

Headline. Overall WER: 24.3 % for our in-house speaker-attributed system; commercial APIs: 43.9 % (ElevenLabs Scribe v2), 51.1 % (Google Chirp 3); best open model: 57.0 % (nvidia/stt_hy_fastconformer_hybrid_large_pc); Whisper large-v3: 88.1 %. On words spoken during overlap: 45.9 % for ours vs 67.3 % for the best commercial API and 83.9 % for the best open model.

Public Armenian ASR benchmarks (Common Voice, FLEURS, ArmBench-ASR) are read or broadcast speech recorded on wideband microphones. Real Armenian phone conversations are different: 8 kHz narrow-band audio after a low-bitrate codec, colloquial Eastern Armenian with Russian and English insertions, and frequent overlapping speech — back-channels, interruptions and both people talking at once. This set measures exactly that.

Every clip is an excerpt of 10–41 s from a two-party customer-service phone call, and every clip contains overlap (at least 0.6 s of both talkers speaking at the same time). Clips were selected by acoustic criteria only (overlap present, cuts at silences where nobody speaks, no word crossing a clip edge) — not by how any recognizer performed on them. The final privacy pass dropped clips in which a recognizer heard personal data (see Privacy); it looked at content, not at accuracy.

Results

All systems receive only the clip (no call context). Open systems were run with default settings (greedy / default generation; Whisper with language="armenian" and 30 s windows). Commercial APIs were called once per clip on 5 November 2026, after the set was fixed, with the released FLAC files as input, the language set to Armenian and all other settings at their defaults: Google Cloud Speech-to-Text v2, model chirp_3, synchronous Recognize, location us, language_codes=["hy-AM"] (Armenian is a Preview language for Chirp 3), automatic decoding, no adaptation, no denoiser; ElevenLabs speech-to-text, model scribe_v2, language_code="hy", tag_audio_events=false. An empty response is scored as an empty transcript (Chirp 3 returned nothing for one clip). Lower is better; 95 % confidence intervals are bootstrapped over clips.

system	WER %	95 % CI	WER on overlapped words %	WER on solo words %	cpWER %
Ours: speaker-attributed, two streams (in-house)	24.27	20.23–28.89	45.92	15.83	24.78
Ours: single stream (in-house)	27.19	23.24–31.98	60.85	15.67	—
ElevenLabs Scribe v2 (API)	43.90	40.01–48.03	67.32	33.55	—
Google Chirp 3 (Cloud Speech-to-Text API)	51.14	45.53–57.42	70.14	43.48	—
nvidia/stt_hy_fastconformer_hybrid_large_pc (VAD segments)	56.99	53.19–61.07	83.94	48.07	—
nvidia/stt_hy_fastconformer_hybrid_large_pc (whole clip)	57.31	53.46–62.01	86.20	47.74	—
facebook omniASR-CTC-1B-v2 (Omnilingual ASR)	76.24	70.63–81.67	92.96	71.04	—
facebook/mms-1b-all (hye adapter)	84.31	80.98–87.60	94.93	80.89	—
openai/whisper-large-v3	88.06	80.95–96.40	90.14	77.36	—
openai/whisper-large-v3-turbo	97.90	89.74–109.10	96.06	88.43	—

WER — one stream: the two-talker reference is serialized four ways (utterances by start time; words by start time; talker 1 then talker 2; talker 2 then talker 1), the two streams of speaker-attributed systems are concatenated in both orders, and per clip the combination with the fewest errors is used, so no system is penalized for the order in which it writes overlapping speech. WER on overlapped words — share of reference words spoken during overlap that the system substituted or missed; WER on solo words — the same for the remaining words. cpWER — for systems that output one transcript per talker (minimum over the speaker permutation). WER can exceed 100 % when a system inserts many words (Whisper often repeats a phrase in a loop on this audio). References spell numbers out in words and scoring does not expand digits, so digits in a hypothesis count as errors: Chirp 3 writes numbers as digits by design (21 of its words on this set); MMS-1b-all has 8 such words, Whisper large-v3 4, the other systems 2 or none.

Fairness notes.

The in-house systems were trained on in-domain Armenian telephone speech, including other calls from the same customer-service line; none of the calls in this test set was used for training or tuning (checked against all training data). The open systems were not trained on narrow-band conversational Armenian, so part of the gap is domain. The training data of the commercial APIs is not public.
The released audio was re-mixed from per-talker channels after voice conversion (see Privacy), which could favour separation-based systems. On an internal control version that keeps the original mono mixture (weaker anonymization, not released) the ranking of the in-house and open systems is the same at the top: both of our systems stay ahead of every open system (ours, speaker-attributed 18.6 %; ours, single stream 23.0 %; nvidia stt_hy (whole clip) 53.8 %; nvidia stt_hy (VAD segments) 55.6 %; omniASR-CTC-1B-v2 68.7 %; MMS-1b-all 77.0 %). The commercial APIs were run on the released audio only.
Voice conversion costs intelligibility: our single-stream system scores 20.33 % WER on the original (non-converted) audio of the same clips and 27.19 % after conversion. Error rates on the original calls are therefore lower than in the table.
On all 97 candidate clips, before the privacy pass removed most of them, the numbers were 29.7 % for our speaker-attributed system and 59.7 % for the best open model: the gap does not come from which clips were kept.

Usage

from datasets import load_dataset

# accept the access conditions on this page first, then log in (huggingface-cli login)
ds = load_dataset("TembraAIPlatform/armenian-telephone-overlap", split="test")
print(ds[0]["id"], ds[0]["text"])

Score a system with eval/score.py from this repository (requires pip install jiwer):

python eval/score.py --meta data/test/metadata.jsonl --hyp my_system.jsonl

The hypothesis file has one JSON object per line: {"id": "hytelovl_0001", "text": "..."} for single-stream systems, or {"id": "hytelovl_0001", "spk": {"a": "...", "b": "..."}} for systems that output one transcript per talker. results.json holds the table above in machine-readable form.

Data fields (`data/test/metadata.jsonl`)

field	meaning
`file_name`, `id`	audio file (FLAC, 8 kHz, 16-bit, mono) and clip id
`duration`	seconds
`text`	reference: both talkers' utterances ordered by start time
`text_spk1`, `text_spk2`	reference per talker
`words`	`{word, start, end, speaker}` for every reference word (seconds; speaker `spk1` / `spk2`)
`muted`	`[start_s, end_s]` spans silenced for privacy; nothing should be transcribed there (empty in this version: clips that would have needed silencing were dropped)
`overlap_sec`, `overlap_events`	total two-talker overlap in the clip and number of overlap regions

Transcripts are in Armenian script as spoken (colloquial forms kept, e.g. էդ, ա for է), lowercase, without punctuation. Scoring normalization: NFKC, lowercase, և → եւ, the in-word marks ՛ ՜ ՞ deleted (ո՞րն → որն, not two words), other punctuation and hesitation fillers removed.

How the references were made

References are per-talker transcripts of the original calls with word-level times, produced by the authors' annotation process. They are not fully human-verified; expect a residual reference error rate of a few percent, concentrated in the hardest overlaps. Words that were silenced for privacy were removed from the references.

Privacy and anonymization

The recordings are real customer-service calls. Before release:

No personal data. Every one of the 97 candidate clips was screened several times, independently, for personal names, organization and product names, phone numbers, identification codes, addresses, e-mails and sensitive personal details: in the reference and in what several recognizers heard in the original and in the released audio. Any clip that contained such an item was dropped. Silencing single words proved unreliable (names are often missing from the reference, and word times around them are imprecise), so no clip in this version has silenced spans; 29 of 97 candidates remain.
Voices replaced. Each talker's channel was converted to a pseudo-voice with kNN-VC (WavLM-Large + HiFi-GAN). A pseudo-voice is built from two Common Voice (CC0) donors, so it does not reproduce any single real person. The pseudo-voice is drawn at random, regardless of the original talker's pitch or gender. Talkers get different pseudo-voices in different calls, which makes it harder (not impossible) to link the same real person across calls; clips from one call share that call's pseudo-voices.
The converted channels were mixed back at their original levels and passed through 8 kHz / 16 kbit/s MP3, like the source audio.

Re-identification test (speaker embeddings, pyannote/wespeaker-voxceleb-resnet34-LM), measured on the 97 candidate clips this set was drawn from (same anonymization; the released subset alone is too small for a stable estimate): top-1 re-identification of the source channel (168 talker channels) among 50 candidate voices: 23 % on original audio vs 5 % after conversion (chance 2.0 %; top-5: 35 % vs 26 %; mean cosine similarity to the true talker 0.67 vs 0.23). Speaking style and tempo are partly preserved.

If you nevertheless recognize a person or organization in this data, please open a discussion in the Community tab of this repository with the clip id only (do not write the name), and the clip will be removed.

Intended use and limitations

Evaluation only. Do not train or fine-tune on this set.
Small (29 clips, 11.1 minutes): report confidence intervals; the bootstrap intervals above are over clips. The set may grow in later versions.
Voice conversion and the mixing of converted channels change the acoustics compared to the original calls; absolute error rates on real calls differ.
One domain (customer service), Eastern Armenian, mostly adult urban speakers.

License and access

CC BY-NC 4.0. Access is gated: by requesting access you agree to use the data for evaluation and research only, not to train or fine-tune models on it, and not to attempt to identify any speaker, person or organization in it.

Citation

@misc{armenian_telephone_overlap_2026,
  title        = {Armenian Telephone Overlap: a test set for overlapped 8 kHz conversational Armenian speech},
  year         = {2026},
  howpublished = {\url{https://huggingface.co/datasets/TembraAIPlatform/armenian-telephone-overlap}}
}

Changes

5 November 2026, update. Added two commercial APIs, Google Chirp 3 and ElevenLabs Scribe v2. The scorer now deletes the in-word Armenian marks ՛ ՜ ՞ instead of splitting the word at them. Of the earlier numbers only nvidia stt_hy changed (57.24 → 56.99 % with VAD segments, 57.69 → 57.31 % on the whole clip); the in-house systems, Whisper, MMS and omniASR are unchanged.

Downloads last month: 1,842

Total file size:

6.37 MB

Access to Armenian Telephone Overlap