debojyoti1986/polycall-bench-corpus

PolyCall Bench — Corpus

The pre-recorded, condition-degraded caller utterances that drive every PolyCall Bench call. PolyCall Bench measures how commercial voice-AI agents degrade under controlled acoustic and network impairment, across real multi-turn phone calls, in five languages. This release is the corpus itself — the scripted caller's audio, ground truth, and seeded degradation parameters — not call recordings or vendor results. A call is produced by playing these files down a live SIP line to a vendor's voice agent and scoring what comes back; that scoring output is a separate dataset.

What's in this dataset

Each row in manifest.jsonl is one scripted caller turn: a short utterance (giving a name, a phone number, or an appointment date) synthesized by TTS, then degraded by one seeded acoustic/network condition and mixed into a continuous ambience bed, exactly as it will be played into a real call. No vendor, no model response, no call outcome — this is the controlled INPUT side only.

Turns	405
Sessions (one bed + its turns)	195
Languages	5 — en-US, en-SG, es-ES, hi-IN, id-ID
Scenarios	S1_phone_only, S2_name_phone, S3_name_phone_date (en-US currently has S3_name_phone_date only - the other two scenarios haven't been built for it yet)
Conditions	15 — ctl_clean, ctl_noise_only, a 6-step cafe-noise SNR ladder (0/5/10/15/20/30 dB), a 4-way ambience sweep (car/office/street/tv at 10 dB), packet loss, jitter, and a bitrate-starved codec

Why a continuous bed, not per-utterance noise

Every turn in a session is cut from one continuous ambience bed, not independently-generated noise per turn — a deliberate design choice. Independently-generated per-turn noise would make the caller appear to teleport between acoustic environments and would destroy any attempt to measure endpointing from the finished audio. bed_path/bed_sha256 in each row point at that session's one shared bed file; offset_s says where in it this turn's speech sits.

Files

• manifest.jsonl — one row per turn, every field self-describing (no config lookup required): lang, scenario, condition_id, the condition's full seeded parameters inlined under condition (never just a name — "cafe noise" is not a condition, the seeded parameter set is), turn_id, field, text (the real ground-truth script, with disfluency/ filler intact) and text_norm (normalized, numbers as digits), typed entities (what a scorer should look for and its exact value), the achieved snr_db_delivered and snr_error_db (how close the real mixed SNR landed to the target — noise mixing is never exact), the per-turn audio_path and sha256, the session's bed_path/bed_sha256, and the seed that produced the whole session.
• audio/<lang>/<scenario>/<condition_id>/ — bed.wav (the continuous ambience bed) and turn_NN.wav (each turn's audio, already mixed into that bed's timeline).

License

Code: MIT. This corpus: CC-BY-4.0.

Citation

This is an early corpus release; a citable version will accompany the full benchmark described in the project's plan.