Datasets:

IOTEverythin
/

voxi-duo-benchmark

Error code:   StreamingRowsError
Exception:    CastError
Message:      Couldn't cast
stoi: double
pesq: double
si_sdr: double
chunks: int64
pitch: double
alignments: null
to
{'alignments': List(List(Json(decode=True)))}
because column names don't match
Traceback:    Traceback (most recent call last):
                File "/src/services/worker/src/worker/utils.py", line 147, in get_rows_or_raise
                  return get_rows(
                      dataset=dataset,
                  ...<4 lines>...
                      column_names=column_names,
                  )
                File "/src/libs/libcommon/src/libcommon/utils.py", line 272, in decorator
                  return func(*args, **kwargs)
                File "/src/services/worker/src/worker/utils.py", line 127, in get_rows
                  rows_plus_one = list(itertools.islice(safe_iter(ds, dataset=dataset), rows_max_number + 1))
                File "/src/services/worker/src/worker/utils.py", line 483, in safe_iter
                  yield from ds.decode(False) if ds.features else ds
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2840, in __iter__
                  for key, example in ex_iterable:
                                      ^^^^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2373, in __iter__
                  for key, pa_table in self._iter_arrow():
                                       ~~~~~~~~~~~~~~~~^^
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2398, in _iter_arrow
                  for key, pa_table in self.ex_iterable._iter_arrow():
                                       ~~~~~~~~~~~~~~~~~~~~~~~~~~~~^^
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 536, in _iter_arrow
                  for key, pa_table in iterator:
                                       ^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 419, in _iter_arrow
                  for key, pa_table in self.generate_tables_fn(**gen_kwags):
                                       ~~~~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/json/json.py", line 343, in _generate_tables
                  self._cast_table(pa_table, json_field_paths=json_field_paths),
                  ~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/json/json.py", line 132, in _cast_table
                  pa_table = table_cast(pa_table, self.info.features.arrow_schema)
                File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2378, in table_cast
                  return cast_table_to_schema(table, schema)
                File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2306, in cast_table_to_schema
                  raise CastError(
                  ...<3 lines>...
                  )
              datasets.table.CastError: Couldn't cast
              stoi: double
              pesq: double
              si_sdr: double
              chunks: int64
              pitch: double
              alignments: null
              to
              {'alignments': List(List(Json(decode=True)))}
              because column names don't match

Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.

Voxi-Duo benchmark

Six models heard the same 20 held-out synthetic recruiter calls (60 s each, callers and scripts not seen in training) and replied in real time, open loop: the recorded caller does not react to the model. Scores come from Whisper large-v3 transcripts of the model channel checked against the call scripts (eval_replies.py); speech quality from torchaudio SQUIM on the same audio. PersonaPlex is nvidia/personaplex-7b-v1 as released, with its stock persona and with a per-call recruiter persona built from the truth file (agent, company, candidate, role, task; never the candidate's answers).

Measure	untouched Moshi	PersonaPlex, stock persona	PersonaPlex + recruiter persona	Voxi-Duo v0.1	Voxi-Duo v0.2	Voxi-Duo v0.3
Calls where it uses task vocabulary	25.0%	40.0%	100.0%	100.0%	100.0%	100.0%
Caller's details said back (of 42)	2.4%	21.4%	33.3%	9.5%	16.7%	14.3%
Share of the call it speaks	15.3%	24.4%	24.5%	27.3%	27.2%	36.7%
Talks over the caller	9.3%	11.1%	10.0%	14.3%	12.3%	20.1%
Reply gap, median	1.85 s	0.43 s	0.4 s	0.9 s	0.75 s	0.8 s
Stops after being interrupted, median	0.75 s	0.35 s	1.08 s	0.93 s	0.25 s	0.7 s
Barge-ins by the model, 20 calls	6	9	14	20	23	34
Words in repetition loops	0.0%	0.0%	0.0%	0.0%	0.0%	0.0%
Speech quality, SQUIM PESQ estimate	3.13	3.53	3.47	2.83	3.02	3.07
STOI	0.99	0.98	0.99	0.94	0.94	0.95
Pitch movement, semitones	3.97	1.9	2.75	3.04	2.96	3.42

report.html is the full write-up with charts; open it in a browser.

Layout

Path	What
`calls/`	The 20 evaluation calls as rendered: left channel scripted agent (TTS), right channel caller. Only the right channel was played to the models.
`truth/`	Per-call ground truth: scenario, agent, candidate details, timing events
`replies/<model>/`	Each model's reply: `.wav` left channel model, right channel caller; `.txt` the text stream the model produced; `.json` Whisper transcript with timestamps; `.prompt.txt` the persona given to PersonaPlex
`scores/`	`eval_replies.py` output per model
`quality/`	SQUIM (PESQ, STOI, SI-SDR) and pitch spread per model

Models: base = kyutai/moshiko-pytorch-bf16 untouched; personaplex-default and personaplex-grounded = nvidia/personaplex-7b-v1 with its stock persona and with a recruiter persona; v0.1..v0.3 = IOTEverythin/voxi-duo-en-in-v2.

Caveats

Open-loop callers (recordings do not react), 20 calls, one seed, synthetic callers, transcript-based scoring. Nothing here measures accent or how natural the voice sounds to a listener.

Licenses

Caller audio: Chatterbox TTS (MIT) from GLOBE reference clips (CC0). Agent side of calls/: Chatterbox from a Svarah clip (CC-BY-4.0). Replies are model outputs: Moshi and Voxi-Duo (CC-BY-4.0); PersonaPlex outputs under the NVIDIA Open Model License.

Downloads last month: 98

Total file size:

684 MB

Models trained or fine-tuned on IOTEverythin/voxi-duo-benchmark

IOTEverythin/voxi-duo-en-in-v2

Audio-to-Audio • Updated 4 days ago