Dataset Viewer
Auto-converted to Parquet Duplicate
Search is not available for this dataset
audio
audioduration (s)
1.45
79.7
End of preview. Expand in Data Studio

mynah-s2s e2e samples

Recorded caller lines for the end-to-end tests of mynah-s2s, a full-duplex voice agent. A test streams these lines into a running server as a microphone would and checks that the agent hears them, calls the right tools and answers every line. With a fixed set of recordings the same audio runs on every machine and every version, and the simulated caller is not the agent's own voice.

All audio is synthetic speech (text-to-speech), short Opus files (mono). Download and use:

make e2e-samples                                   # in the repository: to ~/mynah-s2s/e2e-samples (PART=v2|v3|all)
make test-f2-e2e                                   # Frankenstein v2 uses frankenstein-v2/ when present
make test-f3-e2e                                   # Frankenstein v3 uses frankenstein-v3-african/ when present

Parts

folder engine languages lines voices licence
frankenstein-v2/ Frankenstein v2 (Pocket TTS + per-language delta packages) it, en, fr, de, es, pt, nl, ko, tr, ja, cs, hi, pl, ru, et 81 caller lines + recorded demo dialogues Kyutai Pocket TTS predefined voices, Kyutai voice donations, published voice prompts of the community language models (per language in frankenstein-v2/SOURCES.md) CC-BY-4.0, CC0-1.0, or CC-BY-SA-4.0 where a source is share-alike (Estonian), per language as listed in SOURCES.md
frankenstein-v3-african/ Frankenstein v3 (African profile) sw, sn, lg, zu, xh, ha, ig, yo, rw, am 120 recorded, 110 distinct texts in the manifest (lines whose transcript was too far from the text are listed under rejected) AfricaTTS's own default voice of each language CC-BY-NC-4.0 (non-commercial): output of AfriSpeech/africa-tts, whose weights and voices are CC-BY-NC-4.0

Each part has a manifest.json: {"lines": {text: "lines/<lang>/<scenario>_<i>.ogg"}}. The scenarios are the web UI's demo dialogues: store (a furniture store: price, delivery, goodbye) in every language, and restaurant, weather, travel (it, en, fr in the v2 part; all ten languages in the African part).

The African lines are machine translations of the English lines by AfriqueQwen3.5 9B 50Langs Instruct, kept when their back translation matched the English; a few were fixed by hand. They are not reviewed by native speakers.

How intelligible the lines are

Every line transcribed by Omnilingual ASR CTC v2 3B (bf16, greedy, punctuation and case ignored), WER per language (CER for ja and ko). These are short sentences (3 to 15 per language), so read the numbers as a quality check of the samples, not as ASR benchmark scores.

frankenstein-v3-african (manifest.json keeps only the lines whose transcript is close to the text; the others are listed under rejected with what the ASR heard):

sw sn lg zu xh ha ig yo rw am
.252 .416 .378 .304 .271 .165 .338 .735 .299 .294

frankenstein-v2:

it en fr de es pt nl ko tr ja cs hi pl ru et
.144 .137 .479 .000 .200 .333 .455 .109 .000 .427 .167 .301 .000 .033 .083

The French, Russian and Dutch lines were re-recorded (they were at .77, .52 and .42): Russian is now clean; in French and Dutch the very short lines ("Au nom de Martin.", "Et à Montréal ?") remain the hard ones for this ASR, and part of the WER is digits against words ("500" for "cinq cents"). All three pass the v2 engine's own end-to-end test.

For the African talker itself, the same FLEURS sentences as real recordings and as AfricaTTS speech (in each language's own voice) give a WER within +.03 of real speech in seven of nine languages; Luganda (+.11) and Yoruba (+.12, tones) are its weakest (measured in mynah-s2s, frankenstein2/lab/africa_tts_intelligibility.py).

Licence and use

Mixed, per part, as above: the v2 part follows each voice source's licence (CC-BY-4.0 / CC0-1.0 / CC-BY-SA-4.0); the African part is CC-BY-NC-4.0, for non-commercial use only. No real person's recording is included; every voice is a predefined voice of a TTS model, a voice donated for speech synthesis, or a published voice prompt.

Downloads last month
-