bunsen-tool 0.43.0

bunsen's models on the command line: speech to text (Parakeet, Qwen3-ASR, Voxtral Realtime, Whisper) and who spoke when (Sortformer), text to speech (ZONOS3, Qwen3-TTS), chat (Qwen3, Qwen3.5, Gemma, nanochat) and typed decisions (Laya). The chat driver is also a library, `bunsen_tool::chat`.
Documentation

bunsen-tool

bunsen's models on the command line: speech to text and who spoke when, text to speech, chat, and typed decisions, each loading its models by name from bunsen's catalog.

This README is a tour. Every command and flag documents itself: bunsen-tool --help, bunsen-tool <command> --help, and -h for the short form.

Install

cargo install bunsen-tool --features wgpu     # or vulkan, metal, cuda; the CPU is always there
bunsen-tool asr transcribe recording.wav

On Linux, the microphone and speaker commands link ALSA (libasound2-dev). The default features build every command; --no-default-features --features speech builds asr and tts without audio devices, and --no-default-features alone builds the tool without asr and tts. bunsen-tool devices lists the backends a build has, and what each device does with each float dtype.

Models

bunsen-tool summary lists every model this binary runs, by command, contract and publisher, with its size, dtypes, presets (the default starred) and aliases; each command's list adds licences and what is cached. As it prints them (cargo make cards rebuilds this card):

# bunsen-tool asr ...
# Automated Speech Recognition

# contract: parakeet-asr
# A Parakeet TDT speech recognizer: a Phonon-2 archive, or transformers'
# safetensors and its tokenizer.
# publisher: fermion - FermionResearch, Phonon's authors.

* fermion/phonon-2: [163.5 MB]
  - phonon
  Phonon-2: parakeet-tdt-0.6b-v3 with a five-value encoder, English.

# publisher: nvidia - NVIDIA.

* nvidia/parakeet-tdt-0.6b-v3: [2.51 GB f32]
  parakeet-tdt-0.6b-v3: 600 M parameters, 25 European languages.

# contract: qwen3-asr
# A Qwen3-ASR speech recognizer: a Phonon-1 archive or transformers'
# safetensors, and its tokenizer.
# publisher: fermion - FermionResearch, Phonon's authors.

* fermion/phonon-1: [415.1 MB]
  Phonon-1: Qwen3-ASR-0.6B fine-tuned for English, re-packed.

* fermion/phonon-1-micro: [285.1 MB]
  Phonon-1-Micro: Qwen3-ASR-0.6B fine-tuned for English, re-packed.

* fermion/phonon-1-big: [580.9 MB]
  Phonon-1-Big: Qwen3-ASR-0.6B fine-tuned for English, re-packed.

# publisher: qwen - Qwen, Alibaba Cloud's model team.

* qwen/qwen3-asr-0.6b: [1.58 GB bf16]
  Qwen3-ASR-0.6B: 0.8 B parameters, 30 languages.

* qwen/qwen3-asr-1.7b: [4.09 GB bf16]
  Qwen3-ASR-1.7B: 2.0 B parameters, 30 languages.

# contract: voxtral-asr
# A Voxtral Realtime speech recognizer: its safetensors and its tekken
# vocabulary.
# publisher: mistralai - Mistral AI.

* mistralai/voxtral/voxtral-mini-4b-realtime-2602: [8.87 GB bf16]
  - voxtral-mini-4b-realtime
  Voxtral-Mini-4B-Realtime-2602: 4.4 B parameters, 13 languages, streaming.

# contract: whisper-asr
# An OpenAI Whisper checkpoint, as openai-whisper saves it, and its
# tiktoken vocabulary.
# publisher: openai - OpenAI.

* openai/whisper/large/large-v1: [3.09 GB f16; task: *transcribe|translate]
  1550 M parameters, multilingual.

* openai/whisper/large/large-v2: [3.09 GB f16; task: *transcribe|translate]
  1550 M parameters, multilingual.

* openai/whisper/large/large-v3: [3.09 GB f16; task: *transcribe|translate]
  - large
  1550 M parameters, multilingual, 128 mels.

* openai/whisper/large/large-v3-turbo: [1.62 GB f16]
  - turbo
  809 M parameters, multilingual, 128 mels, four-layer decoder.

* openai/whisper/tiny.en: [76.4 MB f16]
  39 M parameters, English-only.

* openai/whisper/tiny: [76.4 MB f16; task: *transcribe|translate]
  39 M parameters, multilingual.

* openai/whisper/base.en: [146.1 MB f16]
  74 M parameters, English-only.

* openai/whisper/base: [146.1 MB f16; task: *transcribe|translate]
  74 M parameters, multilingual.

* openai/whisper/small.en: [484.5 MB f16]
  244 M parameters, English-only.

* openai/whisper/small: [484.4 MB f16; task: *transcribe|translate]
  244 M parameters, multilingual.

* openai/whisper/medium.en: [1.53 GB f16]
  769 M parameters, English-only.

* openai/whisper/medium: [1.53 GB f16; task: *transcribe|translate]
  769 M parameters, multilingual.

# contract: sortformer-diarizer
# A Sortformer speaker diarizer: a `.nemo` archive, which carries its
# config, or transformers' safetensors.
# publisher: nvidia - NVIDIA.

* nvidia/sortformer/nemotron-3-diarization: [397.0 MB f32; latency: ultra-low|very-low|*low|offline; postprocess: *none|dihard3|callhome]
  Nemotron-3-Diarization: 100 M parameters, up to 8 speakers, 10 ms frames,
  streaming.

* nvidia/sortformer/diar_streaming_sortformer_4spk-v2: [471.4 MB f32; latency: ultra-low|*low|high|offline; postprocess: *none|dihard3|callhome]
  - sortformer-v2
  Streaming Sortformer v2: 117 M parameters, up to 4 speakers, 80 ms
  frames.

* nvidia/sortformer/diar_streaming_sortformer_4spk-v2.1: [471.4 MB f32; latency: ultra-low|*low|high|offline; postprocess: *none|dihard3|callhome]
  - sortformer-v2.1
  Streaming Sortformer v2.1: 117 M parameters, up to 4 speakers, 80 ms
  frames.

# contract: silero-vad
# The Silero voice-activity detector at 16 kHz: upstream's state dict.
# publisher: silero - Silero, snakers4/silero-vad's authors.

* silero/silero_vad/vad-16k: [1.2 MB f32; vad-policy: *silero|faster-whisper]
  Silero VAD, the 16 kHz model, from upstream's safetensors.

# bunsen-tool tts ...
# Text to Speech

# contract: zonos3-tts
# A ZONOS3 text-to-speech model as Zyphra's transformers packaging lays it
# out: the language model and the codec decoder in one safetensors file,
# their geometry from `config.json`.
# publisher: zyphra - Zyphra.

* zyphra/zonos3/zonos3-mini: [2.39 GB bf16 f32]
  ZONOS3-Mini: 800 M parameters (100 M active), 44.1 kHz speech in 9
  languages, 27 stock voices.

# contract: qwen3-tts
# A Qwen3-TTS model as Qwen's repositories lay it out: the talker and its
# code predictor in one safetensors file, the 12 Hz codec in another, their
# geometry from their `config.json`s; the tokenizer as its slow files.
# publisher: qwen - Qwen, Alibaba Cloud's model team.

* qwen/qwen3-tts/qwen3-tts-12hz-0.6b-customvoice: [2.50 GB f32]
  Qwen3-TTS-12Hz-0.6B-CustomVoice: 9 stock voices in 10 languages and two
  Chinese dialects, 24 kHz speech; no style instruction.

* qwen/qwen3-tts/qwen3-tts-12hz-1.7b-customvoice: [4.52 GB f32]
  Qwen3-TTS-12Hz-1.7B-CustomVoice: 9 stock voices in 10 languages and two
  Chinese dialects, 24 kHz speech, steered by a style instruction.

* qwen/qwen3-tts/qwen3-tts-12hz-1.7b-voicedesign: [4.52 GB f32]
  Qwen3-TTS-12Hz-1.7B-VoiceDesign: a voice from a description, in 10
  languages, 24 kHz speech; no stock voices.

# bunsen-tool chat ...
# Chat and Text Generation

# contract: qwen3-lm
# A dense Qwen3 causal LM, as transformers lays it out: its geometry from
# `config.json`.
# publisher: qwen - Qwen, Alibaba Cloud's model team.

* qwen/qwen3/qwen3-0.6b: [1.51 GB bf16; thinking: *thinking|no-thinking]
  Qwen3-0.6B: thinking and non-thinking chat.

* qwen/qwen3/qwen3-0.6b-base: [1.20 GB bf16; thinking: no-thinking]
  Qwen3-0.6B-Base: the pretrained model, no chat tuning.

* qwen/qwen3/qwen3-1.7b: [4.07 GB bf16; thinking: *thinking|no-thinking]
  Qwen3-1.7B: thinking and non-thinking chat.

* qwen/qwen3/qwen3-1.7b-base: [3.45 GB bf16; thinking: no-thinking]
  Qwen3-1.7B-Base: the pretrained model, no chat tuning.

* qwen/qwen3/qwen3-4b: [8.06 GB bf16; thinking: *thinking|no-thinking]
  Qwen3-4B: thinking and non-thinking chat.

* qwen/qwen3/qwen3-4b-base: [8.05 GB bf16; thinking: no-thinking]
  Qwen3-4B-Base: the pretrained model, no chat tuning.

* qwen/qwen3/qwen3-4b-instruct-2507: [8.06 GB bf16; thinking: no-thinking]
  Qwen3-4B-Instruct-2507: non-thinking chat, 256K context.

* qwen/qwen3/qwen3-4b-thinking-2507: [8.06 GB bf16; thinking: thinking]
  Qwen3-4B-Thinking-2507: thinking chat, 256K context.

* qwen/qwen3/qwen3-8b: [16.39 GB bf16; thinking: *thinking|no-thinking]
  Qwen3-8B: thinking and non-thinking chat.

* qwen/qwen3/qwen3-8b-base: [16.39 GB bf16; thinking: no-thinking]
  Qwen3-8B-Base: the pretrained model, no chat tuning.

* qwen/qwen3/qwen3-4b-saferl: [8.06 GB bf16; thinking: *thinking|no-thinking]
  Qwen3-4B-SafeRL: Qwen3-4B tuned for safety by reinforcement learning,
  thinking and non-thinking chat.

* qwen/qwen3/qwen3guard-gen-0.6b: [1.51 GB bf16; thinking: no-thinking]
  Qwen3Guard-Gen-0.6B: safety moderation, rating a prompt or a reply Safe,
  Unsafe or Controversial, with its categories.

* qwen/qwen3/qwen3guard-gen-4b: [8.83 GB bf16; thinking: no-thinking]
  Qwen3Guard-Gen-4B: safety moderation, rating a prompt or a reply Safe,
  Unsafe or Controversial, with its categories.

* qwen/qwen3/qwen3guard-gen-8b: [16.39 GB bf16; thinking: no-thinking]
  Qwen3Guard-Gen-8B: safety moderation, rating a prompt or a reply Safe,
  Unsafe or Controversial, with its categories.

# contract: qwen3.5-lm
# A dense Qwen3.5 checkpoint's text model, as transformers lays it out: its
# geometry from `config.json`.
# publisher: qwen - Qwen, Alibaba Cloud's model team.

* qwen/qwen3.5/qwen3.5-0.8b: [1.76 GB bf16; thinking: thinking|*no-thinking]
  Qwen3.5-0.8B: thinking and non-thinking chat (its text model).

* qwen/qwen3.5/qwen3.5-0.8b-base: [1.76 GB bf16; thinking: no-thinking]
  Qwen3.5-0.8B-Base: the pretrained model, no chat tuning (its text model).

* qwen/qwen3.5/qwen3.5-2b: [4.56 GB bf16; thinking: thinking|*no-thinking]
  Qwen3.5-2B: thinking and non-thinking chat (its text model).

* qwen/qwen3.5/qwen3.5-2b-base: [4.56 GB bf16; thinking: no-thinking]
  Qwen3.5-2B-Base: the pretrained model, no chat tuning (its text model).

* qwen/qwen3.5/qwen3.5-4b: [9.33 GB bf16; thinking: *thinking|no-thinking]
  Qwen3.5-4B: thinking and non-thinking chat (its text model).

* qwen/qwen3.5/qwen3.5-4b-base: [9.33 GB bf16; thinking: no-thinking]
  Qwen3.5-4B-Base: the pretrained model, no chat tuning (its text model).

* qwen/qwen3.5/qwen3.5-9b: [19.32 GB bf16; thinking: *thinking|no-thinking]
  Qwen3.5-9B: thinking and non-thinking chat (its text model).

* qwen/qwen3.5/qwen3.5-9b-base: [19.32 GB bf16; thinking: no-thinking]
  Qwen3.5-9B-Base: the pretrained model, no chat tuning (its text model).

* qwen/qwen3.5/qwen3.5-27b: [55.58 GB bf16; thinking: *thinking|no-thinking]
  Qwen3.5-27B: thinking and non-thinking chat (its text model); 54 GB in
  bf16, more than a card holds.

# contract: gemma-lm
# A Gemma 1, 2 or 3 text model, as transformers lays it out: its geometry
# from `config.json`.
# publisher: gemma - Google's Gemma models, from their unsloth/ mirrors.

* gemma/gemma/gemma-3-270m: [570.8 MB bf16; thinking: no-thinking]
  Gemma 3 270M: the pretrained model.

* gemma/gemma/gemma-3-270m-it: [570.8 MB bf16; thinking: no-thinking]
  Gemma 3 270M, instruction-tuned.

* gemma/gemma/gemma-3-1b-pt: [2.03 GB bf16; thinking: no-thinking]
  Gemma 3 1B: the pretrained model.

* gemma/gemma/gemma-3-1b-it: [2.03 GB bf16; thinking: no-thinking]
  Gemma 3 1B, instruction-tuned.

* gemma/gemma/gemma-3-4b-pt: [8.63 GB bf16; thinking: no-thinking]
  Gemma 3 4B: the pretrained model's text model.

* gemma/gemma/gemma-3-4b-it: [8.63 GB bf16; thinking: no-thinking]
  Gemma 3 4B, instruction-tuned: its text model.

* gemma/gemma/gemma-2-2b: [5.25 GB bf16; thinking: no-thinking]
  Gemma 2 2B: the pretrained model, stored in f32.

* gemma/gemma/gemma-2-2b-it: [5.25 GB bf16; thinking: no-thinking]
  Gemma 2 2B, instruction-tuned.

* gemma/gemma/gemma-2-9b: [18.52 GB bf16; thinking: no-thinking]
  Gemma 2 9B: the pretrained model.

* gemma/gemma/gemma-2-9b-it: [18.50 GB bf16; thinking: no-thinking]
  Gemma 2 9B, instruction-tuned.

* gemma/gemma/gemma-2b: [5.03 GB bf16; thinking: no-thinking]
  Gemma 2B (Gemma 1): the pretrained model.

* gemma/gemma/gemma-2b-it: [5.03 GB bf16; thinking: no-thinking]
  Gemma 2B (Gemma 1), instruction-tuned.

* gemma/gemma/gemma-7b: [17.11 GB bf16; thinking: no-thinking]
  Gemma 7B (Gemma 1): the pretrained model.

* gemma/gemma/gemma-7b-it: [17.09 GB bf16; thinking: no-thinking]
  Gemma 7B (Gemma 1), instruction-tuned.

# contract: nanochat-lm
# A nanochat GPT, as upstream's training run writes it: its geometry from
# the run's `meta_*.json`.
# publisher: nanochat-students - nanochat-students, who publish nanochat training runs.

* nanochat-students/base-d20: [2.08 GB bf16; thinking: no-thinking]
  - nanochat-d20-base
  nanochat-students' d20 after pretraining (step 21400): upstream's 2025
  model, 561M, base.

* nanochat-students/mid-d20: [2.08 GB bf16; thinking: no-thinking]
  - nanochat-d20-mid
  nanochat-students' d20 after midtraining (step 765): upstream's 2025
  model, 561M, chat format.

* nanochat-students/rl-d20: [2.08 GB bf16; thinking: no-thinking]
  - nanochat-d20-rl
  nanochat-students' d20 after its RL stage (step 650): upstream's 2025
  model, 561M, chat.

# publisher: nitishpandey04 - nitishpandey04, who publishes nanochat training runs.

* nitishpandey04/nanochat-d18-base: [2.05 GB; thinking: no-thinking]
  nitishpandey04's d18 after pretraining (step 3712): upstream's 2026
  model, 702M, base.

* nitishpandey04/nanochat-d18-sft: [2.05 GB; thinking: no-thinking]
  nitishpandey04's d18 after SFT (step 478): upstream's 2026 model, 702M,
  chat.

# bunsen-tool decide ...
# Typed Decisions

# contract: laya-decision
# A Laya decision model: a ModernBERT encoder and a decision head, in one
# safetensors file.
# publisher: laya - convaiinnovations, Laya's authors.

* laya/laya/laya: [846.2 MB f16]
  - laya-en
  Laya, English: ModernBERT-large (28 layers, 1024 wide), 421M parameters,
  prompts of 512 tokens.

* laya/laya/laya-multilingual: [678.2 MB f16]
  Laya, multilingual: mmBERT-base (22 layers, 768 wide, a 256k vocabulary),
  322M parameters, prompts of 1024 tokens.

* laya/laya/laya-typed-decisions: [846.2 MB f16]
  Laya, typed decisions: the English model trained on typed questions,
  prompts of 1024 tokens.

Models are fetched into the weights cache on first use. fetch NAME... brings them in ahead of time, for a deployment that runs --offline.

asr: speech to text

Name a preset, a model chosen by what it is for, and asr picks the family, the model and its settings, for files (transcribe) or the microphone (live):

bunsen-tool asr transcribe recording.wav                       # English, the default preset
bunsen-tool asr transcribe --preset multi --json recording.wav # any language, as JSON lines
bunsen-tool asr live --preset translate-en                     # the microphone, translated to English
bunsen-tool asr transcribe --language de --quality fast recording.wav
bunsen-tool asr presets list                                   # what each preset is for, and runs
preset for transcribe runs live runs
en (default) English Parakeet, fermion/phonon-2 the same
multi many languages, found from the speech Qwen3-ASR, qwen/qwen3-asr-1.7b qwen/qwen3-asr-0.6b
translate-en other speech, as English text Whisper, large-v3[translate] tiny[translate]

A preset chooses its model by the catalog's tags, so a better model, when it arrives, is chosen without the command line changing. --speech, --task, --quality and --language choose a preset, or adjust one, by what is needed:

command line runs task
(nothing) en: Parakeet transcribe
--language de multi: Qwen3-ASR transcribe
--task translate translate-en: Whisper translate
--speech en --task translate en: Parakeet (English is never translated to English) transcribe
--preset multi --task translate multi, translating: Whisper translate
--preset en --language de an error: en hears only English

presets show prints what a preset runs as the asr family command line, and -v logs the choice.

Model families, directly

asr family <family> runs one family, with only the flags that apply to it: parakeet, qwen3-asr, voxtral and whisper, each with list, fetch, transcribe and live.

bunsen-tool asr family qwen3-asr list
bunsen-tool asr family whisper transcribe --model 'openai/small[translate]{beam_size: 1}' recording.wav
bunsen-tool asr family parakeet live --emission drafting

--model takes a lookup expression: a name, its presets in brackets, and a YAML patch of its settings in braces (quote it: [ and ] glob). list shows each model's presets. The settings a patch can set:

family settings
parakeet window_seconds (60)
qwen3-asr window_seconds (30)
voxtral window_seconds (30); delay_ms (480): how far the transcript runs behind the audio
whisper language, task, beam_size (5), timestamps, max_tokens, condition_on_previous_text, fallback

Every family also takes hf:org/repo, a Hugging Face repo in transformers' layout. --emission says when to decode: offline (whole windows), conservative (each speech region as it ends, found by a VAD) or drafting (conservative, plus a draft of the speech so far, which live shows).

Who spoke when

bunsen-tool asr diarize run --rttm meeting.wav > meeting.rttm            # speaker turns alone
bunsen-tool asr transcribe --preset multi --diarize --json meeting.wav   # each statement with its speaker
bunsen-tool asr live --diarize

diarize runs NVIDIA's Sortformer diarizers: nvidia/nemotron-3-diarization (up to 8 speakers) by default, or Sortformer v2 and v2.1 (up to 4). Their presets choose the latency (ultra-low to offline) and the post-processing.

Where the transcript goes

Without --publish, to stdout, as text, --json lines or --rttm. --publish CONFIG sends it to every target a config names at once, inline or from a file, in YAML or JSON:

times: true
targets:
  - type: stdout                    # text: [00:00:01.200 --> 00:00:03.400] [speaker_0] hello there
  - type: file
    path: meeting.jsonl
    format: jsonl                   # text (the default), jsonl, rttm (needs --diarize)
    append: false
  - type: log                       # a JSON line per record, at `level` under `target`
    level: info
    target: bunsen_tool::transcript
  - type: webhook                   # each record POSTed as JSON (the `webhook` feature)
    url: https://example.com/hook
    headers: {Authorization: Bearer TOKEN}
    events: [statement]
  - type: mqtt                      # each record published to {prefix}/{event} (the `mqtt` feature)
    broker: mqtt://broker.lan:1883  # plain TCP; 1883 by default
    prefix: bunsen-tool/room-1
    qos: 1                          # 0, 1 (the default) or 2
    retain: false
    client_id: bunsen-tool-room-1    # bunsen-tool-{pid} by default
    username: asr                   # and password, both optional

The records are statement (committed text, with its speaker under --diarize), draft (live's latest draft) and transcript (a file's whole transcript); times, speakers and events may be set for every target or per target. A network target delivers from its own thread, so a slow one never holds the audio up.

tts: text to speech

bunsen-tool tts run -o hello.wav "Hello, this is ZONOS3."
bunsen-tool tts run --play "Spoken, not saved."
bunsen-tool tts run --model qwen3-tts --voice ryan -o ryan.wav "Hello from Qwen."
bunsen-tool tts voices                                   # a model's stock voices

ZONOS3-Mini is the default; Qwen3-TTS's checkpoints add a style instruction (--instruct) and a voice described in words (--describe). The text is normalized as the model's release normalizes it, so $12.50 is spoken as "twelve dollars fifty cents", and long text is spoken as several takes. A take is sampled; --seed repeats one.

chat: Qwen3, Qwen3.5, Gemma and nanochat

bunsen-tool chat run --no-think "What is the capital of France?"
bunsen-tool chat repl --model qwen3-4b-instruct-2507 --stats
bunsen-tool chat run --model gemma-3-270m-it "What is the capital of France?"
bunsen-tool chat family nanochat run "What is the capital of France?"

run answers one prompt; repl holds a conversation, prefilling only what the KV cache does not already hold. The chat driver is also a library, bunsen_tool::chat.

decide: typed questions answered by Laya

decide run answers questions about a context with calibrated probabilities over each question's options, in one forward pass of a Laya decision model. The request is YAML or JSON:

context: "Customer: my kettle arrived with a cracked lid and I asked for a refund twice with no answer."
choices:
  - type: choice                    # a pick among named options
    question: "What should the agent do next?"
    options: [refund, replace, {escalate: "hand the case to a supervisor"}, close]
  - type: noul                      # a yes/no probability; `score` is a level on a scale
    question: "Is the customer satisfied?"
bunsen-tool decide run request.yml

Each result has each option's probability, the most probable (answer) and a confidence.

text: normalization on its own

$ printf 'It costs $5.32\non 3/4/2025.\n' | bunsen-tool text normalize --normalize zyphra --normalize transcript --language en
it costs five dollars thirty two cents on march fourth twenty twenty five

The normalizers TTS speaks through and word error rates compare through; text normalizers lists them.

Choosing a model

chat, decide and tts choose their model by a pipeline over the catalog's builds, by their tags:

bunsen-tool chat run --model qwen --le parameters 2e9 "Hi"                       # a group, filtered
bunsen-tool chat run --model qwen/qwen3 --eq thinking never --asc parameters "Hi"  # stages, in order

--model names a build, an alias or a group; stage flags (--keep, --le, --weigh, --asc, ...) keep and order within it; presets (--preset, --on, --off) apply to the one chosen. -v logs the choice and why. bunsen-tool catalog browses the catalog itself: tree, show, resolve, and select, the same pipeline.

The weights cache

bunsen-tool cache path                          # where it is
bunsen-tool cache list [NAME]                   # what it holds, and when each file was last used
bunsen-tool cache clear [NAME] --age 30d --yes  # what has not been used for 30 days

Every fetched file is pinned to a SHA-256 and verified as it lands. --cache-dir moves the cache, and --offline never reaches the network. Every command also takes --device, and the commands that load a model --min-precision f32, a floor under the dtypes its weights are placed in.