baptistefrancois1/tool-calling-s2s-v2

Tool-calling speech-to-speech dialogues (v2)

Synthetic spoken dialogues between a visitor and the voice receptionist of a fictional company, in French and English, with the tool calls the assistant makes between turns. Built to finetune a speech-to-speech model (LFM2-Audio) so that it answers by voice and decides when to call a tool.

Four tools exist: find_employee, notify_employee_via_teams, call_backup_staff, get_company_info. A dialogue is a list of turns: the visitor's speech, the assistant's decision at each point (a call, or none), the tool results, and the assistant's spoken answer. Every employee, visitor, email (@example-sia.test), building and wifi password is fictional.

Repository: baptistefrancois1/tool-calling-s2s-v2.

Contents

Language	Dialogues	Spoken clips	Hours of audio	Audio (GB)
fr	1075	3300	4.3	0.74
en	1075	3300	5.25	0.91

The French and the English files share their dialogue identifiers: the same scenario is written and spoken once in each language. Voices (Voxtral built-in voices), clips per voice:

Language	Role	Voices
fr	assistant	fr_female (1650)
fr	visitor	fr_female (836), fr_male (814)
en	assistant	casual_female (1650)
en	visitor	casual_male (422), cheerful_female (404), neutral_female (412), neutral_male (412)

One assistant voice per language, so a model trained on it never has to guess which voice to answer in. In French, fr_female also speaks the visitor in about half of the dialogues (only fr_male and fr_female speak French without an accent).

Files

• dataset/tool_calling_v2_fr.jsonl, dataset/tool_calling_v2_en.jsonl: one dialogue per line.
• dataset/clips_tool_calling_v2_fr.jsonl, dataset/clips_tool_calling_v2_en.jsonl: one row per clip (text, voice, duration).
• audio/tool_calling_v2/fr/<shard>/<clip>.wav, audio/tool_calling_v2/en/<shard>/<clip>.wav: 24 kHz mono WAV. The audio field of a turn is the path of its clip inside this repository.

Fields of a dialogue: id (tc2-NNNNN), lang, family, tools_declared (the tools listed in its system prompt), turns, meta. Fields of a turn: role (user, assistant, tool), type (spoken, tool_call, no_call, tool_result), text, audio, voice, calls, results, slot, hints. A no_call turn is the assistant's explicit decision that no tool is needed; it precedes every spoken answer.

meta records what the generator knew about the scenario: the kind (off_topic_kind, f5_branch, f6_branch, f7_branch, chitchat_kind, info_type, follow_up, badge_context), visitor, visitor_intro (self: the visitor says their name; asked: the assistant asks for it), employee_id, and source_id (the identifier in the generator's folder). Held-out evaluation dialogues can be stratified on it.

Scenario families (per language)

Family	Dialogues	What happens
F1_no_call	150	small talk, and thanks after a finished notification: no tool is called
F2_off_scope	150	requests the receptionist has no tool for (ten kinds): refused, no call
F3_info	200	company information (wifi, directions, floors, name), some with a follow-up answered from the result already in context
F4_chain	200	an exact meeting: find the employee, notify them
F5_confirm	150	an inexact search: confirm first, then notify, or retry after a wrong person, or call the reception staff after a second failure
F6_fallback	150	a name not found (asked again once, then found or the staff is called) and a notification that failed
F7_recover	75	an out-of-scope request refused, then a real request (information or meeting) in the same dialogue

A visitor's name is always heard before it is used: either said in the visitor's first line, or asked by the assistant before the first call that needs it. Names the visitor says wrongly (Fontaine heard as Fantaine) are searched as heard, and the confirmation uses the directory's spelling.

How it was made

1. Structure by code: the family, the tool calls and their arguments, the tool results and the decision turns are generated deterministically from a fictional directory (150 employees, 320 visitor names, three building variants). Nothing the assistant calls is invented by a model.
2. Text by an LLM (claude-sonnet-5, same prompts in both languages): only the spoken lines are written, under checks (no leaked language, a refusal offers nothing, every name copied into a call has been heard).
3. Speech by Voxtral-4B-TTS-2603, built-in voices only.
4. Cadence: so that the assistant's answers can be laid out in an interleaved text/audio format (6 text tokens for 12 audio frames), a short silence (median 0.24 s, at most 0.75 s) was appended to the French assistant clips whose audio ended before their text.

Limits

• Everything is synthetic: the speech is clean, the dialogues are short (1 to 3 exchanges), and the visitors are cooperative. It does not replace real speech.
• About a quarter of the French visitor lines carry a gender agreement ("désolée", "pressé") while the voice is drawn independently of the text.
• The French and English files are translations of the same scenarios, not independent corpora.
• The directory has 150 employees: a model finetuned on it may memorise their names.

License

cc-by-nc-4.0. The audio was generated with Voxtral-4B-TTS-2603, whose reference voices are licensed CC BY-NC 4.0 and whose license the model inherits; check its model card before any commercial use. The text was written by a language model from deterministic skeletons.