Datasets:
Maya voice agent: spoken test sets
The three test sets of synthetic spoken phone calls from the paper Teaching an Open Speech-to-Speech Model to Take Bookings (link to follow), with the caller audio. Each conversation lists the tools the agent has, the caller turns, and the reference agent turns with their tool calls and mock results.
| Files | Test | Conversations | Tool turns |
|---|---|---|---|
eval.jsonl + general_eval.answered.jsonl |
Round 1 test: the travel agent, plus general questions | 343 + 69 | 315 |
r2_eval.jsonl |
Round 2 test: held-out cities, two held-out clinics, 10 emergency calls | 329 | 403 |
biz_eval.jsonl |
A car service centre, a business never seen in training | 50 | 57 |
audio/<id>_u<k>.wav is caller turn k of conversation id: 16 kHz mono, voiced by Kokoro-82M and
Indic Parler-TTS. All voices are synthetic. The rows of general_eval.answered.jsonl with an empty
answer were dropped from the test and have no audio.
To score a model, put the .jsonl files in train/data/ and audio/ in train/audio/ of the
code repository, then run
python evaluate.py --test v1|v2|newbiz --model <llm.gguf> --out results/x.jsonl.
- Downloads last month
- -