Expand description
§bunsen-tool
bunsen-tool is bunsen’s models on the command line: bunsen-tool asr
(speech to text, and who spoke when), bunsen-tool chat (chat with a
Qwen3, Qwen3.5, Gemma or nanochat checkpoint), bunsen-tool decide
(typed questions about a context, answered by a Laya decision model),
bunsen-tool text (text normalization on its own), bunsen-tool cache
(the weights cache), bunsen-tool summary and bunsen-tool devices. Its
README is the binary’s guide.
The library is all of it, one module per command, each with what the
command runs and its command tree, nested as the commands are: a
{group}_command_group per group, a {group}_{command} per command, every
type named by its command path (DecideRunCommand, ChatKitCommandGroup)
and re-exported at the command’s module:
asr(featurespeech): presets and model families for speech to text, the stream drivers, the publishers;asr::AsrCommandGroup.tts(featurespeech): text to speech with ZONOS3-Mini or Qwen3-TTS;tts::TtsCommandGroup.chat: the chat driver;chat::ChatCommandGroup.decide: Laya decisions on typed questions;decide::DecideCommandGroup.text: text normalization on its own;text::TextCommandGroup.summaryanddevices: the builds this binary runs, in brief, by command, contract and publisher; the backends and their devices.support: what the command trees share.
Modules§
- asr
- Speech to text, and who spoke when, as
bunsen-tool asrruns them. - chat
- Chat with bunsen’s Qwen3 (
bunsen::kits::language::generation::qwen3), Gemma (bunsen::kits::language::generation::gemma) or nanochat (bunsen::kits::language::generation::nanochat): what turns a conversation into the ids the model reads and its ids back into text, and the driver that holds a conversation on the KV cache. - decide
- Typed questions about a context, answered with calibrated probabilities by
a Laya decision model (
bunsen::kits::language::decision::laya), asbunsen-tool decideruns them. - devices
bunsen-tool devices: what this build can run on, and what each device does with each float dtype.- summary
bunsen-tool summary: the builds this binary can run, in brief: by command, contract and publisher.- support
- What the command trees share: a lookup expression with presets or a patch
added, a row’s presets and a checkpoint’s row as the listings print them
(
print_model_row), the--presetflags, the--formatchoice (Format), an ordered document for it to render (Value), and “did you mean?” on a failed lookup (did_you_mean); withlive, the host’s audio devices (find_audio_device,list_audio_devices). - text
- Text normalization on its own, as
bunsen-tool textruns it: the normalizers a TTS driver or a word error rate reads text through (bunsen::kits::support::text). - tts
- Text to speech, as
bunsen-tool ttsruns it: every TTS family’s checkpoints (TtsFamily: ZONOS3-Mini and Qwen3-TTS) through the shared TTS driver.