Full-Duplex-Bench: A Benchmark to Evaluate Full-duplex Spoken Dialogue Models on Turn-taking Capabilities
Paper • 2503.04721 • Published • 4
Request access to the Full-Duplex-Bench v1.0 model outputs
Requests are reviewed manually; expect a short delay.
This dataset contains MODEL OUTPUTS produced by running third-party and our own full-duplex speech models through the Full-Duplex-Bench v1.0 protocol (CC BY-NC 4.0). It does NOT contain the benchmark's stimulus audio (CANDOR and ICC recordings of real people, and FDB's synthetic clips) — obtain those from the benchmark authors. By requesting access you agree to use this data for non-commercial research only, to cite Full-Duplex-Bench and the stimulus corpora, and to respect the terms of the model providers (NVIDIA PersonaPlex, Sakana AI KAME / Kyutai Moshi, KRAFTON Raon-SpeechChat).
Log in or Sign Up to review the conditions and access this dataset content.
Full-Duplex-Bench v1.0 model outputs (fdb-v1-outputs-v1)
7997 model responses = 11 benchmark runs × the 727 stimuli of Full-Duplex-Bench v1.0 (pause handling, backchannel, smooth turn-taking, user interruption). The runs cover 7 systems; several differ only in voice prompt, prompting regime or weights, which is the point — those are controlled pairs. For every stimulus and run you get the model's own reply channel as lossless FLAC — time-synchronous with the stimulus, so turn-taking events can be measured against it — plus the ASR transcript of that reply, the model's own text stream, the user-interruption judge rating where one was made, the harness's per-run result files, and a leaderboard. The stimuli are not included (see Rejoining the stimuli).
This is a research archive of benchmark runs by the JSALT 2026 Conversational AI Simulator
workstream (WP1), produced with the fdb-v1-repro harness (commit f166c53-dirty).
Metric definitions are the benchmark's; our numbers are not the paper's numbers (different voices,
seeds, and a local judge — see Caveats).
Runs
Each run is a datasets config — load_dataset("MagicLuke/fdb-v1-outputs-v1", "<run>") gives that run's
727 rows. Each model run is its own config subset (load_dataset("MagicLuke/fdb-v1-outputs-v1", "<run>")); there is no aggregate config.
| config (run) | system under test | voice | prompt regime | seed | oracle | generated |
|---|---|---|---|---|---|---|
personaplex_natf2 |
PersonaPlex-7B (nvidia/personaplex-7b-v1) |
NATF2.pt (packaged voice prompt) | FDB category personas: 'You enjoy having a good conversation.' (pause / backchannel / turn-taking), 'You are a wise and friendly teacher...' (user interruption) | 42424242 | — | 2026-07-17 |
personaplex_varf0 |
PersonaPlex-7B (nvidia/personaplex-7b-v1) |
VARF0.pt (packaged voice prompt; the paper-closest, chattier voice) | FDB category personas (same as NATF2) | 42424242 | — | 2026-07-21 |
kame_shim |
KAME (Sakana AI; Moshi S2S + asynchronous oracle LLM) (SakanaAI/kame) |
KAME's own (no voice prompt) | bare: no text prompt; KAME's released oracle system prompt applies to every stimulus | — | unsloth/gemma-4-E4B-it served locally through an OpenAI-compatible shim under the name gpt-4.1; trigger ASR nvidia/parakeet-tdt-0.6b-v2 on the same shim | 2026-07-21 |
kame_oracle |
KAME (Sakana AI; Moshi S2S + asynchronous oracle LLM) (SakanaAI/kame) |
KAME's own (no voice prompt) | bare: no text prompt; KAME's released oracle system prompt applies to every stimulus | — | openai/gpt-4.1-nano through an OpenAI-compatible institutional gateway; trigger ASR nvidia/parakeet-tdt-0.6b-v2 on the local shim | 2026-07-21 |
raon |
Raon-SpeechChat-9B (KRAFTON) (KRAFTON/Raon-SpeechChat-9B) |
spk_ref.wav (Raon's speaker-embedding reference clip) | bare: Raon's own scaffold only ('You are engaging in real-time conversation.'), listen-first (speak_first=False) | 42 | — | 2026-08-22 |
oif_stage2_337_base |
PersonaPlex-7B + our online-instruction-following stage-2 adapter, merged (MagicLuke/personaplex-oif-ins-token-v1 (stage2/checkpoint_000337 LoRA, folded at scaling 2.0 into nvidia/personaplex-7b-v1)) |
NATF2.pt (packaged voice prompt — the BASE model's conditions) | FDB category personas (same as personaplex_natf2) | 42424242 | — | 2026-08-21 |
brainoff |
PersonaPlex-7B + our stage-2 adapter, merged, primed as duplex-brain's caller seat with the brain off (MagicLuke/personaplex-oif-ins-token-v1 (stage2/checkpoint_000337 LoRA, folded at scaling 2.0 into nvidia/personaplex-7b-v1)) |
stage2_caller_voice.wav — the synthetic caller speaker the adapter was trained under (stage-2 manifest row 8, ex Correction.001.k8), fed raw (no loudness normalization) | one fixed training card for every stimulus: the row-8 persona card ('You are Olivia ... calling Express Railways ... Work through these goals strictly in order ...'), i.e. a CALLER persona against user-speech stimuli | 42424242 | — | 2026-08-22 |
s2_varf0 |
PersonaPlex-7B + our online-instruction-following stage-2 adapter, merged (MagicLuke/personaplex-oif-ins-token-v1 (stage2/checkpoint_000337 LoRA, folded at scaling 2.0 into nvidia/personaplex-7b-v1)) |
VARF0.pt (packaged voice prompt; the paper-closest, chattier voice) | FDB category personas (same as personaplex_natf2) | 42424242 | — | 2026-08-24 |
gdpo_v2v600_varf0 |
PersonaPlex-7B + our GDPO-v2 RL adapter (run gdpo-v2f-r4, checkpoint v600), merged (gdpo-v2f-r4 v600 (RL LoRA on nvidia/personaplex-7b-v1, folded at scaling 2.0; adapter not yet released)) |
VARF0.pt (packaged voice prompt; the paper-closest, chattier voice) | FDB category personas (same as personaplex_varf0 / s2_varf0) | 42424242 | — | 2026-08-25 |
gdpo1200_brainoff |
PersonaPlex-7B + our GDPO v1200 RL adapter, merged, primed as duplex-brain's caller seat with the brain off (gdpo-duplex adapter v1200 (RL LoRA r128/32) folded at scaling 2.0 into nvidia/personaplex-7b-v1 (extended vocab dq8/instr32002); adapter not yet released) |
stage2_caller_voice.wav (raw, no LUFS) | one fixed training customer card (cards/brainoff_row8.txt) for pause/turn-taking/backchannel; the FDB standard assistant persona ('a wise and friendly teacher') for user-interruption, so that category is comparable to the other rows | 42424242 | — | 2026-08-23 (4 categories) / 2026-08-24 (user-interruption, assistant persona) |
gdpo_simple_v2f_v600_varf0 |
PersonaPlex-7B + our GDPO simple-reward RL adapter (checkpoint v600), merged (gdpo-simple-v2f v600 (RL LoRA on nvidia/personaplex-7b-v1, folded at scaling 2.0; adapter not yet released)) |
VARF0.pt (packaged voice prompt; the paper-closest, chattier voice) | FDB category personas (same as personaplex_varf0 / s2_varf0 / gdpo_v2v600_varf0) | 42424242 | — | 2026-08-27 |
personaplex_natf2— The harness baseline. Numbers move strongly with the voice prompt (see the VARF0 row and the probe in fdb-v1-repro/README.md).personaplex_varf0— Same model, same seed, same prompts as personaplex_natf2 — only the voice prompt differs; this is the row that matches the PersonaPlex paper's Table 2 regime.kame_shim— Not the paper's configuration (paper: hosted GPT-4.1 oracle, Google STT). Streamed in real time over KAME's WebSocket server; ~0.1 s opus priming skew is a known caveat on latency metrics.kame_oracle— Same as kame_shim with a real OpenAI oracle (gpt-4.1-nano, not the paper's gpt-4.1).raon— In-process, one 80 ms output frame per input frame (frame-locked, no transport skew). Sampling temp 0.9 / top-k 66 / top-p 0.95 (Raon's duplex_infer.yaml), eager attention. Per-sample generation stats in results/raon/raon_stats/.oif_stage2_337_base— Our finetuned checkpoint scored under the base model's voice and prompts; it goes quiet on this benchmark (smooth-turn TOR 0.24). Compare with brainoff, the training-matched priming of the same weights.brainoff— Training-matched priming of the same weights as oif_stage2_337_base: plain persona , no instruction tokens, no brain. Expect leading/chatty behaviour by construction (it is a caller, not a responder).s2_varf0— THE headline row for our finetune, and the controlled pair of personaplex_varf0: same voice, same prompts, same seed, same harness -- only the weights differ, so the delta is the finetune's. Winner of a 5-cell voice x prompt probe (fdb-v1-repro scripts/stage2_probe.sh): beats our KAME rows on 8/10 columns; vs the base at identical conditioning it is more disciplined (pause TOR 0.153/0.245 vs 0.642/0.718, backchannel TOR 0.018 vs 0.418, smooth-turn latency 0.007 vs 0.020) but less proactive (smooth-turn TOR 0.756 vs 0.983, interruption rating 4.04 vs 4.68, interruption latency 0.859 vs 0.214) -- the instruction-following/initiative trade, isolated from any conditioning confound.gdpo_v2v600_varf0— The RL-v2 checkpoint under the same conditioning as s2_varf0 and personaplex_varf0, so the three form a controlled ladder: base -> stage-2 SFT -> GDPO-v2 RL, one voice (VARF0), one prompt regime (FDB category personas), one seed. GDPO v2 (adaptive-KL controller, beta 0.05 / target 0.3) ran to step 1181 and v600 is its best held-out checkpoint (compliance@0.5 .579, fits .906, no-instruction delta +.398); folded with scaling 2.0, adapter sha256 1f1581a7...305d2. Brain off: FDB-v1 feeds one recorded stimulus to one model, so there is no second seat for the steering LLM to act on.gdpo1200_brainoff— RL checkpoint gdpo-duplex/adapters/v1200 (step 1200; val c_goal 0.706). The 4 timing categories are persona-agnostic; user-interruption was regenerated with the assistant persona because the fixed customer card is the wrong role for a QA task.gdpo_simple_v2f_v600_varf0— Same harness, voice, prompts and seed as personaplex_varf0, s2_varf0 and gdpo_v2v600_varf0, so the four form a controlled ladder and this row differs from gdpo_v2v600_varf0 by the RL reward alone. The simple reward replaces the checklist c_goal + per-turn fits + strict-seat judge passes with ONE judge call per sample asking 3-level instruction coverage and one naturalness/own-seat question, at the same 33/13/54 instruction/naturalness/turn-taking variance balance. Against gdpo_v2v600_varf0 it improves backchannel TOR (0.218 vs 0.327) and smooth-turn latency (0.156 vs 0.203) and loses on pause-handling TOR (0.394/0.500 vs 0.365/0.417).
Leaderboard
Rendered from the exported results/ files (leaderboard.md holds the same table). * = FDB paper
Table III (arXiv 2503.04721); ‡ = PersonaPlex paper Table 2 (arXiv 2602.06053); † = our
User-Interruption rating comes from a local judge (unsloth/gemma-4-E4B-it served as
gpt-4-turbo), not OpenAI — comparable across our rows, not with the paper rows.
| Model | Pause Handling | Backchannel | Smooth Turn Taking | User Interruption | ||||||
|---|---|---|---|---|---|---|---|---|---|---|
| Synthetic | Candor | ICC | Candor | Synthetic | ||||||
| TOR (↓) | TOR (↓) | TOR (↓) | Freq (↑) | JSD (↓) | TOR (↑) | Latency (↓) | TOR (↑) | GPT-4o (↑) | Latency (↓) | |
| dGSLM* | 0.934 | 0.935 | 0.691 | 0.015 | 0.934 | 0.975 | 0.352 | 0.917 | 0.201 | 2.531 |
| Moshi* | 0.985 | 0.980 | 1.000 | 0.001 | 0.957 | 0.941 | 0.265 | 1.000 | 0.765 | 0.257 |
| Freeze-Omni* | 0.642 | 0.481 | 0.636 | 0.001 | 0.997 | 0.336 | 0.953 | 0.867 | 3.615 | 1.409 |
| Gemini Live* | 0.255 | 0.310 | 0.091 | 0.012 | 0.896 | 0.655 | 1.301 | 0.891 | 3.376 | 1.183 |
| PersonaPlex‡ | 0.584 | 0.662 | 0.327 | 0.025 | 0.649 | 0.992 | 0.070 | 1.000 | 4.210 | 0.400 |
| Qwen-2.5-Omni‡ | – | – | – | – | – | – | – | – | 4.590 | 2.740 |
| PersonaPlex (NATF2) | 0.234 | 0.343 | 0.200 | 0.098 | 0.755 | 0.958 | 0.401 | 0.955 | 4.785† | 0.192 |
| PersonaPlex (VARF0) | 0.642 | 0.718 | 0.418 | 0.176 | 0.700 | 0.983 | 0.020 | 0.935 | 4.684† | 0.214 |
| KAME (local oracle shim) | 0.620 | 0.718 | 0.727 | 0.025 | 0.897 | 0.975 | 0.151 | 0.820 | 3.521† | 0.356 |
| KAME (gpt-4.1-nano oracle) | 0.613 | 0.685 | 0.709 | 0.028 | 0.893 | 0.966 | 0.205 | 0.845 | 3.083† | 0.441 |
| Raon-SpeechChat-9B | 0.650 | 0.880 | 0.473 | 0.171 | 0.694 | 1.000 | 0.017 | 1.000 | 3.791† | 0.236 |
| OIF stage2-337 (NATF2, FDB prompts) | 0.000 | 0.009 | 0.255 | 0.045 | 0.786 | 0.244 | 0.997 | 0.930 | 3.914† | 0.977 |
| duplex-brain (brain off, training card+voice) | 0.124 | 0.236 | 0.073 | 0.011 | 0.953 | 0.857 | 0.992 | 0.960 | 2.312† | 0.908 |
| duplex-brain (brain off, VARF0, FDB prompts) | 0.153 | 0.245 | 0.018 | 0.056 | 0.795 | 0.756 | 0.007 | 0.950 | 4.037† | 0.859 |
| PersonaPlex GDPO v2 v600 (brain off, VARF0, FDB prompts) | 0.365 | 0.417 | 0.327 | 0.118 | 0.724 | 1.000 | 0.203 | 1.000 | 4.725† | 0.268 |
| PersonaPlex GDPO v1200 (brain off; interruption uses the FDB assistant persona) | 0.971 | 0.991 | 0.691 | 0.062 | 0.830 | 1.000 | 0.000 | 1.000 | 4.650† | 0.283 |
| PersonaPlex GDPO simple-v2f v600 (brain off, VARF0, FDB prompts) | 0.394 | 0.500 | 0.218 | 0.113 | 0.750 | 1.000 | 0.156 | 1.000 | 4.790† | 0.264 |
Rejoining the stimuli
data/<run>/<category>/<id>/ mirrors FDB v1.0's own <category>/<id>/ directory names exactly
(numeric ids are the upstream folder names). Download the v1.0 stimuli from the benchmark authors
(link in v1_v1.5/dataset/README.md of the FDB repository), then pair each of our output.flac
with the stimulus input.wav of the same <category>/<id>; the two have identical durations
(our harness pads/trims the reply to the stimulus length). Categories and sizes:
| category | n | what is measured | stimulus source |
|---|---|---|---|
synthetic_pause_handling |
137 | pause handling (TOR during the user's pauses, lower is better) | synthetic: GPT-4o text rendered with ChatTTS, by the FDB authors |
candor_pause_handling |
216 | pause handling (TOR, lower is better) | CANDOR corpus excerpts (Reece et al. 2023), real human speech |
candor_turn_taking |
119 | smooth turn-taking (TOR higher is better, latency lower is better) | CANDOR corpus excerpts, real human speech |
icc_backchannel |
55 | backchannel (TOR lower, frequency higher, JSD vs human timing lower) | In Conversation Corpus (Umair, Sarathy & de Ruiter, Findings EMNLP 2024), real human speech |
synthetic_user_interruption |
200 | user interruption (TOR higher, judge rating 1-5 higher, latency lower) | synthetic: GPT-4o text rendered with ChatTTS, by the FDB authors |
Layout
metadata/<run>.jsonl the same rows for one run — the per-run configs
data/<run>/<category>/<id>/output.flac the model's reply channel (24 kHz mono, lossless)
data/<run>/<category>/<id>/output.json nvidia/parakeet-tdt-0.6b-v2 ASR of output.flac: {text, chunks:[{text, timestamp:[s,e]}]}
data/<run>/<category>/<id>/model_text.json the model's own text stream (a JSON string; not ASR)
data/<run>/<category>/<id>/rating.json user-interruption judge {analysis, rating 1-5} (only samples with TOR=1 are rated)
results/<run>/{pause_synthetic,pause_candor,smooth_turn_taking,backchannel,user_interruption}.txt
the FDB evaluation scripts' output (per-sample lines + aggregates)
results/<run>/summary.md the run's metric table (regenerated from the txt files)
results/raon/raon_stats/shard_*.jsonl Raon generation stats per sample (frames, fps, rms, text chars)
leaderboard.md the table above
provenance/runs.json, build_info.json this release's spec, harness commit, per-run counts + metrics
metadata.jsonl columns: run, label, model, model_id, category, sample_id, stimulus (<category>/<id>), duration_s, sample_rate, audio, asr, model_text, rating (int|null), rating_path, st_tor, st_latency_s (smooth-turn per-sample TOR / raw latency, candor_turn_taking only; raw latencies can be negative — the aggregate clamps them at 0), voice, prompt_regime, seed, oracle.
from datasets import load_dataset
from huggingface_hub import snapshot_download
import soundfile as sf
rows = load_dataset("MagicLuke/fdb-v1-outputs-v1", "raon", split="train") # one run's rows
# load_dataset("MagicLuke/fdb-v1-outputs-v1", "personaplex_natf2", split="train") # one run subset (727 rows)
root = snapshot_download("MagicLuke/fdb-v1-outputs-v1", repo_type="dataset",
allow_patterns=["data/raon/*", "results/raon/*"]) # that run's files
r = next(r for r in rows if r["category"] == "candor_turn_taking")
reply, sr = sf.read(f"{root}/{r['audio']}") # pair with the FDB stimulus r["stimulus"] + "/input.wav"
FLAC is bit-exact with the harness's WAV output (6.77 GB → 1.59 GB; replies are mostly silence while the user speaks).
How it was generated
- Generation — fdb-v1-repro: gen_fdb_v1.py (PersonaPlex family, in-process moshi offline loop), gen_fdb_v1_kame.py (KAME WebSocket server, real-time 80 ms opus frames), gen_fdb_v1_raon.py (Raon in-process, one 80 ms frame per input frame).
- ASR — nvidia/parakeet-tdt-0.6b-v2 (upstream FDB get_transcript/asr.py) -> output.json.
- Scoring — upstream FDB v1_v1.5/evaluation scripts; user-interruption rating by a LOCAL judge (unsloth/gemma-4-E4B-it served OpenAI-compatible under the name gpt-4-turbo), not OpenAI.
- Output format — output.wav = the model's own reply channel, 24 kHz mono PCM_16, time-synchronous with (same duration as) the stimulus; stored here as lossless FLAC.
- Per-run voice / prompt / seed / oracle are in the Runs table and
provenance/runs.json; generation dates are the output files' dates on the cluster (NCSA Delta).
Caveats
- Voice prompts move these numbers a lot. Swapping PersonaPlex's packaged voice NATF2 → VARF0 (same model, prompts, seed) changes pause TOR 0.23 → 0.64 and smooth-turn latency 0.40 → 0.02 s. Compare systems only at a stated voice + prompt; the two PersonaPlex rows bracket this.
- The user-interruption rating is from a local judge (gemma-4-E4B-it served as
gpt-4-turbo), not the paper's GPT-4o/GPT-4-turbo;rating.jsonexists only where the model took the turn. - KAME's oracle is substituted: a local gemma-4 shim under the name
gpt-4.1(kame_shim) orgpt-4.1-nanoover a gateway (kame_oracle), with Parakeet as the trigger ASR — not the paper's hosted GPT-4.1 + Google STT. KAME was streamed in real time over its WebSocket server (~0.1 s opus priming skew on latencies). - Raon is run bare and listen-first; it is the most eager speaker here (barges in on most pauses) and the fastest responder.
- Our checkpoint rows:
oif_stage2_337_basescores our finetuned PersonaPlex under the base model's voice/prompts (it goes quiet);brainoffprimes the same weights the way they were trained — a caller persona card + the training speaker's voice — so it leads the conversation by construction. - Raw per-sample smooth-turn latencies in
results/*/smooth_turn_taking.txtandst_latency_scan be negative (model spoke before the user's turn end); the FDB aggregate clamps them at 0. - No confidence intervals; one stimulus is ~0.5–1.8 pp of a category.
Licensing and attribution — non-commercial
This dataset contains model outputs only. The Full-Duplex-Bench v1.0 stimuli are not
redistributed: the candor_* categories are excerpts of the CANDOR corpus and icc_backchannel
of the In Conversation Corpus — recordings of real people available from their authors on request —
and FDB releases them under CC BY-NC 4.0 with the instruction to respect those upstream terms;
the synthetic_* stimuli were generated by the FDB authors (GPT-4o text, ChatTTS; MIT). Each
output.flac is the model's own synthetic reply channel, produced while the stimulus was streamed
as the user channel; it contains no stimulus audio. Because the outputs exist only by running the
FDB v1.0 protocol, we treat this release conservatively as a Full-Duplex-Bench derivative and
license it CC BY-NC 4.0 (non-commercial). All speech here is synthetic model output.
Cite the benchmark and the stimulus corpora: Lin et al., Full-Duplex-Bench: A Benchmark to Evaluate Full-Duplex Spoken Dialogue Models on Turn-taking Capabilities, arXiv:2503.04721 (2025); Reece et al., The CANDOR corpus, Science Advances 9(13), 2023; Umair, Sarathy & de Ruiter, Large Language Models Know What To Say But Not When To Speak, Findings of EMNLP 2024.
Model terms also apply to the generated audio: NVIDIA's license for nvidia/personaplex-7b-v1
(including its packaged voices NATF2/VARF0), Sakana AI's terms for SakanaAI/kame and Kyutai's
for Moshi, KRAFTON's CC BY-NC 4.0 for KRAFTON/Raon-SpeechChat-9B; our stage-2 adapter
(MagicLuke/personaplex-oif-ins-token-v1) is a PersonaPlex derivative. ASR by
nvidia/parakeet-tdt-0.6b-v2; ratings by unsloth/gemma-4-E4B-it. Access is gated so these terms
are acknowledged before download.
Citation
If you use these outputs, cite Full-Duplex-Bench (above) and the JSALT 2026 Conversational AI
Simulator project (citation to follow in the fdb-v1-repro README).
Changelog
- 2026-08-27: added gdpo_simple_v2f_v600_varf0 (GDPO simple-reward v600
- Downloads last month
- 27