neudecide 0.0.1__py3-none-any.whl

neudecide/__init__.py

@@ -0,0 +1,4 @@
+from .model import NeuDecide
+
+__version__ = "0.0.1"
+__all__ = ["NeuDecide"]

neudecide/audio.py

@@ -0,0 +1,58 @@
+"""Audio input: WAV loading with the standard library and resampling with soxr."""
+
+import wave
+from pathlib import Path
+
+import numpy as np
+import soxr
+
+
+def load_audio(path):
+    """Reads a PCM WAV file. Returns (mono float32 waveform in [-1, 1], sample rate)."""
+    try:
+        with wave.open(str(path), "rb") as f:
+            channels, width, sr = f.getnchannels(), f.getsampwidth(), f.getframerate()
+            raw = f.readframes(f.getnframes())
+    except wave.Error as e:
+        raise ValueError(
+            f"{path}: can only read PCM WAV files ({e}); load other formats yourself and "
+            "pass a numpy array with sample_rate="
+        ) from e
+    if width == 1:
+        x = (np.frombuffer(raw, dtype=np.uint8).astype(np.float32) - 128.0) / 128.0
+    elif width == 2:
+        x = np.frombuffer(raw, dtype="<i2").astype(np.float32) / 32768.0
+    elif width == 3:
+        b = np.frombuffer(raw, dtype=np.uint8).reshape(-1, 3).astype(np.int32)
+        v = b[:, 0] | (b[:, 1] << 8) | (b[:, 2] << 16)
+        x = (np.where(v >= 1 << 23, v - (1 << 24), v)).astype(np.float32) / float(1 << 23)
+    elif width == 4:
+        x = np.frombuffer(raw, dtype="<i4").astype(np.float32) / float(1 << 31)
+    else:
+        raise ValueError(f"{path}: unsupported sample width {width}")
+    return x.reshape(-1, channels).mean(axis=1), sr
+
+
+def resample(x, orig_sr, target_sr):
+    """Band-limited resampling with soxr (its default "HQ" quality)."""
+    x = np.asarray(x, dtype=np.float32)
+    if orig_sr == target_sr or x.size == 0:
+        return x
+    return soxr.resample(x, orig_sr, target_sr)
+
+
+def prepare_audio(audio, sample_rate, target_sr):
+    """A path or array -> mono float32 at `target_sr`. Arrays are (samples,) or
+    (samples, channels) and default to `target_sr` when `sample_rate` is None."""
+    if isinstance(audio, (str, Path)):
+        audio, sample_rate = load_audio(audio)
+    else:
+        audio = np.asarray(audio, dtype=np.float32)
+        if audio.ndim == 2:
+            audio = audio.mean(axis=1)
+        elif audio.ndim != 1:
+            raise ValueError(
+                f"audio must be 1-D (samples,) or 2-D (samples, channels), got {audio.shape}"
+            )
+        sample_rate = sample_rate or target_sr
+    return resample(audio, sample_rate, target_sr)

neudecide/grammar.py

@@ -0,0 +1,154 @@
+"""Grammar-constrained decoding for tool calls, with llguidance.
+
+The model writes a JSON list of calls. `tool_call_grammar` describes exactly
+that shape, specialised to the tools offered:
+
+    [ ]  |  [ CALL (, CALL)* ]      CALL := {"name": NAME, "arguments": {ARGS}}
+
+NAME must be one of the tool names, and ARGS may only use that tool's
+parameter names as keys (in any order, at most as many keys as it has
+parameters), each with any well-formed JSON value. Whitespace between JSON
+tokens is free, so compact and `json.dumps`-spaced output are both accepted.
+
+Decoding policy lives in a `Selector`: anything with `select(logits) -> id`.
+`GreedySelector` is a plain argmax; `ConstrainedSelector` masks the logits to
+the tokens the grammar allows next and takes the argmax of what's left.
+"""
+
+import json
+import re
+
+import numpy as np
+from llguidance import LLMatcher, LLTokenizer, TokenizerWrapper
+from llguidance.numpy import apply_token_bitmask_inplace
+
+WS = "▁"  # SentencePiece's whitespace marker
+
+
+def tool_parameters(tools):
+    """{tool name: [argument keys]} for a list (or JSON string) of tool
+    definitions, each with a `name` and `parameters` either as JSON Schema
+    (`{"type": "object", "properties": {...}}`) or as a flat `{key: {...}}` mapping."""
+    if isinstance(tools, str):
+        tools = json.loads(tools)
+    params = {}
+    for tool in tools or []:
+        if not isinstance(tool, dict) or "name" not in tool:
+            continue
+        p = tool.get("parameters") or {}
+        # a flat mapping may itself have a parameter called "type", so check the schema shape
+        is_schema = p.get("type") == "object" and isinstance(p.get("properties"), dict)
+        props = p["properties"] if is_schema else p
+        params[tool["name"]] = list(props) if isinstance(props, dict) else []
+    return params
+
+
+def tool_call_grammar(tools):
+    """A Lark grammar, in llguidance's dialect, for the answer to `tools`."""
+    calls = []
+    for name, keys in tool_parameters(tools).items():
+        # patternProperties, not properties: llguidance emits properties in
+        # schema order, and the model needn't. maxProperties bounds repeated
+        # keys, which a JSON grammar can't rule out.
+        arguments = {
+            "type": "object",
+            "patternProperties": {f"^{re.escape(k)}$": {} for k in keys},
+            "additionalProperties": False,
+            "maxProperties": len(keys),
+        }
+        calls.append(
+            {
+                "type": "object",
+                "properties": {"name": {"const": name}, "arguments": arguments},
+                "required": ["name", "arguments"],
+                "additionalProperties": False,
+            }
+        )
+    schema = {
+        "type": "array",
+        "items": {"anyOf": calls} if calls else False,
+        # \uXXXX escapes in values, as json.dumps writes non-ASCII text by default
+        "x-guidance": {"json_allow_general_unicode_escapes": True},
+    }
+    # leading whitespace: SentencePiece's dummy prefix puts a space before the "["
+    return "start: WS? calls\n" r"WS: /[ \t\n\r]+/" "\n" f"calls: %json {json.dumps(schema)}"
+
+
+class _Vocab:
+    """The attributes llguidance.TokenizerWrapper reads off a tokenizer."""
+
+    def __init__(self, tokens, special_token_ids, eos_token_id, encode):
+        self.tokens = tokens
+        self.special_token_ids = special_token_ids
+        self.eos_token_id = eos_token_id
+        self.bos_token_id = None
+        self._encode = encode
+
+    def __call__(self, text):
+        return self._encode(text)
+
+
+def ll_tokenizer(tokenizer, eos_id, special_ids=()):
+    """The SentencePiece vocabulary as an llguidance tokenizer: each piece as
+    the bytes it adds to the output (▁ -> space, <0xNN> -> that byte).
+    Control / unknown / unused pieces and `special_ids` are special tokens,
+    which the grammar never produces."""
+    sp = tokenizer.sp
+    tokens, special = [], set(special_ids)
+    for i in range(sp.vocab_size()):
+        piece = sp.id_to_piece(i)
+        if sp.is_byte(i):
+            tokens.append(bytes([int(piece[1:-1], 16)]))
+        else:
+            tokens.append(piece.replace(WS, " ").encode("utf-8"))
+        if sp.is_control(i) or sp.is_unknown(i) or sp.is_unused(i):
+            special.add(i)
+    vocab = _Vocab(tokens, sorted(special), eos_id, tokenizer.sp_fragment.encode)
+    return LLTokenizer(TokenizerWrapper(vocab))
+
+
+def tool_call_matcher(ll_tok, tools):
+    """A fresh llguidance matcher for the answer to `tools`."""
+    matcher = LLMatcher(ll_tok, tool_call_grammar(tools), log_level=0)
+    if matcher.is_error():
+        raise ValueError(f"couldn't build the tool-call grammar: {matcher.get_error()}")
+    return matcher
+
+
+# --------------------------------------------------------------------------
+# Token selection
+# --------------------------------------------------------------------------
+
+
+class GreedySelector:
+    """Plain argmax."""
+
+    def select(self, logits):
+        return int(np.argmax(logits))
+
+
+class ConstrainedSelector:
+    """Greedy decoding restricted to `matcher`'s grammar, for one sequence.
+
+    The first step is forced to `tool_call_id` (every answer starts with it)
+    and `eos_id` is forced as soon as the grammar is complete."""
+
+    def __init__(self, matcher, tool_call_id, eos_id):
+        self.matcher = matcher
+        self.tool_call_id = tool_call_id
+        self.eos_id = eos_id
+        self.started = False
+
+    def select(self, logits):
+        if not self.started:
+            self.started = True
+            return self.tool_call_id
+        if self.matcher.is_accepting():
+            return self.eos_id
+        logits = np.array(logits, dtype=np.float32).reshape(1, -1)
+        mask = np.frombuffer(self.matcher.compute_bitmask(), dtype=np.int32)
+        apply_token_bitmask_inplace(logits, mask)
+        token = int(np.argmax(logits))
+        if not self.matcher.consume_token(token):
+            return self.eos_id  # nothing fits; shouldn't happen
+        return token

neudecide/model.py

@@ -0,0 +1,204 @@
+import json
+import logging
+from pathlib import Path
+
+import numpy as np
+import onnxruntime as ort
+
+from .audio import prepare_audio
+from .grammar import ConstrainedSelector, GreedySelector, ll_tokenizer, tool_call_matcher
+from .tokenizer import Tokenizer
+
+logger = logging.getLogger(__name__)
+
+DEFAULT_REPO = "neuphonic/neudecide"
+DEFAULT_VARIANT = "q4"  # int4 MatMulNBits graphs, the ones the native runtime ships
+TOKENIZER_FILES = ("tokenizer.model", "tokenizer.json")  # first one present wins
+
+
+class NeuDecide:
+    """Speech + a list of tools -> tool calls, with onnxruntime.
+
+    Three ONNX graphs run per request:
+      audio_encoder   once per utterance:            waveform -> audio frames
+      tool_encoder    once per (utterance, tools):   -> cross-attention keys / values
+      decoder_step    once per output token, with a growing self-attention KV cache
+
+        model = NeuDecide.from_pretrained()
+        calls = model.generate("command.wav", tools)
+        # [{"name": "set_timer", "arguments": {"minutes": 5}}]
+    """
+
+    def __init__(self, model_dir, variant=DEFAULT_VARIANT, threads=None, providers=None):
+        """model_dir: a directory with config.json, tokenizer.model (or tokenizer.json)
+        and the ONNX graphs.
+        variant: which set of graphs from config.json's "variants".
+        threads: onnxruntime intra-op threads (None = onnxruntime's default)."""
+        model_dir = Path(model_dir)
+        self.config = json.loads((model_dir / "config.json").read_text())
+        tokenizer = next((model_dir / f for f in TOKENIZER_FILES if (model_dir / f).exists()), None)
+        if tokenizer is None:
+            raise FileNotFoundError(f"{model_dir} has no {' or '.join(TOKENIZER_FILES)}")
+        self.tokenizer = Tokenizer.from_file(tokenizer)
+        self.special = self.config["special_tokens"]
+        self.sample_rate = self.config["sample_rate"]
+        self.variant = variant
+        graphs = self.graph_files(self.config, variant)
+
+        options = ort.SessionOptions()
+        if threads:
+            options.intra_op_num_threads = threads
+        providers = list(providers or ["CPUExecutionProvider"])
+        self.audio_encoder, self.tool_encoder, self.decoder_step = (
+            ort.InferenceSession(
+                str(model_dir / graphs[k]), sess_options=options, providers=providers
+            )
+            for k in ("audio_encoder", "tool_encoder", "decoder_step")
+        )
+        self._ll_tokenizer = None
+
+    @staticmethod
+    def graph_files(config, variant):
+        variants = config.get("variants") or {"fp32": {"graphs": config["graphs"]}}
+        if variant not in variants:
+            raise ValueError(f"model has no {variant!r} variant (has: {sorted(variants)})")
+        return variants[variant]["graphs"]
+
+    @classmethod
+    def from_pretrained(
+        cls, repo_id=DEFAULT_REPO, variant=DEFAULT_VARIANT, revision=None, token=None, **kwargs
+    ):
+        """Downloads (or reuses the cached) model from the Hugging Face Hub."""
+        from huggingface_hub import hf_hub_download
+        from huggingface_hub.utils import EntryNotFoundError
+
+        def fetch(name):
+            return hf_hub_download(repo_id, name, revision=revision, token=token)
+
+        config_path = Path(fetch("config.json"))
+        config = json.loads(config_path.read_text())
+        for name in TOKENIZER_FILES:
+            try:
+                fetch(name)
+                break
+            except EntryNotFoundError:
+                continue
+        for name in cls.graph_files(config, variant).values():
+            fetch(name)
+        return cls(config_path.parent, variant=variant, **kwargs)
+
+    # -- stages ---------------------------------------------------------------
+
+    def encode_tools(self, tools):
+        """[<tools>] + the tokenized tool list, serialized with json.dumps defaults
+        (as in training) and truncated to max_tools_len."""
+        text = tools if isinstance(tools, str) else json.dumps(tools)
+        ids = [self.special["tools"]] + self.tokenizer.encode(text)
+        limit = self.config["max_tools_len"]
+        if len(ids) > limit:
+            logger.warning("tool list is %d tokens; truncated to %d", len(ids), limit)
+            ids = ids[:limit]
+        return np.asarray([ids], dtype=np.int64)
+
+    def encode_audio(self, waveform):
+        """16 kHz mono waveform -> (audio_embeds, audio_len)."""
+        waveform = np.asarray(waveform, dtype=np.float32)
+        waveform = waveform[: int(self.config["max_audio_seconds"] * self.sample_rate)]
+        return self.audio_encoder.run(
+            None,
+            {
+                "waveform": waveform[None, :],
+                "waveform_len": np.asarray([waveform.shape[0]], dtype=np.int64),
+            },
+        )
+
+    def selector(self, tools, constrained=True):
+        """The token-selection policy for one generation."""
+        if not constrained:
+            return GreedySelector()
+        if self._ll_tokenizer is None:
+            self._ll_tokenizer = ll_tokenizer(
+                self.tokenizer, self.special["eos"], self.special.values()
+            )
+        matcher = tool_call_matcher(self._ll_tokenizer, tools)
+        return ConstrainedSelector(matcher, self.special["tool_call"], self.special["eos"])
+
+    def generate_ids(
+        self, audio, tools, sample_rate=None, constrained=True, max_new_tokens=None, selector=None
+    ):
+        """Output token ids (EOS excluded). `selector` overrides the decoding
+        policy: any object with `select(logits) -> token id`."""
+        waveform = prepare_audio(audio, sample_rate, self.sample_rate)
+        audio_embeds, audio_len = self.encode_audio(waveform)
+        cross_key, cross_value, encoder_mask = self.tool_encoder.run(
+            None,
+            {
+                "audio_embeds": audio_embeds,
+                "audio_len": audio_len,
+                "tools_ids": self.encode_tools(tools),
+            },
+        )
+
+        selector = selector or self.selector(tools, constrained)
+        c = self.config
+        past_key = past_value = np.zeros(
+            (c["num_decoder_layers"], 1, c["num_kv_heads"], 0, c["head_dim"]), dtype=np.float32
+        )
+        token, out = self.special["eos"], []  # the decoder starts from EOS
+        for _ in range(max_new_tokens or c["max_answer_len"]):
+            logits, past_key, past_value = self.decoder_step.run(
+                None,
+                {
+                    "input_ids": np.asarray([[token]], dtype=np.int64),
+                    "past_key": past_key,
+                    "past_value": past_value,
+                    "cross_key": cross_key,
+                    "cross_value": cross_value,
+                    "encoder_mask": encoder_mask,
+                },
+            )
+            token = selector.select(logits[0])
+            if token == self.special["eos"]:
+                break
+            out.append(token)
+        return out
+
+    def decode(self, ids):
+        """Token ids -> answer text, without the structural tokens."""
+        ids = list(ids)
+        if ids and ids[0] == self.special["tool_call"]:
+            ids = ids[1:]
+        structural = {self.special[k] for k in ("pad", "eos", "bos", "tool_call", "tools")}
+        return self.tokenizer.decode([i for i in ids if i not in structural])
+
+    # -- high level -----------------------------------------------------------
+
+    def generate_text(self, audio, tools, **kwargs):
+        """The raw answer text: a JSON list of calls."""
+        return self.decode(self.generate_ids(audio, tools, **kwargs))
+
+    def generate(self, audio, tools, **kwargs):
+        """Tool calls for an utterance: a list of {"name", "arguments"} dicts,
+        [] when no tool applies.
+
+        audio: a path to a WAV file, or a numpy array (samples,) or
+            (samples, channels) -- pass sample_rate= if it isn't 16 kHz.
+        tools: a list of tool definitions (name, description, JSON-Schema
+            parameters), or the same as a JSON string.
+        constrained: restrict output to valid calls of the given tools (default True).
+        max_new_tokens: output length limit (default: the model's max_answer_len).
+        """
+        text = self.generate_text(audio, tools, **kwargs)
+        return parse_calls(text)
+
+
+def parse_calls(text):
+    """A JSON list of objects with a "name", else [] (logged)."""
+    try:
+        calls = json.loads(text)
+    except json.JSONDecodeError:
+        calls = None
+    if isinstance(calls, list) and all(isinstance(c, dict) and "name" in c for c in calls):
+        return calls
+    logger.warning("model output isn't a list of tool calls: %r", text)
+    return []

neudecide/tokenizer.py

@@ -0,0 +1,61 @@
+"""SentencePiece tokenizer, from the model's tokenizer.model or tokenizer.json.
+
+tokenizer.model is SentencePiece's own file (a serialized ModelProto).
+tokenizer.json holds the same BPE model as JSON: its pieces (text, score, type)
+plus the normalizer flags encoding depends on; it's turned back into a
+ModelProto here so the sentencepiece library does the encoding and decoding.
+"""
+
+import json
+from pathlib import Path
+
+import sentencepiece as spm
+from sentencepiece import sentencepiece_model_pb2 as sp_pb2
+
+
+def model_proto(spec):
+    """tokenizer.json contents -> a SentencePiece ModelProto (identity normalizer)."""
+    if spec.get("model_type") != "bpe":
+        raise ValueError(f"unsupported tokenizer model_type {spec.get('model_type')!r}")
+    proto = sp_pb2.ModelProto()
+    proto.trainer_spec.model_type = sp_pb2.TrainerSpec.BPE
+    proto.trainer_spec.unk_id = spec["unk_id"]
+    proto.trainer_spec.byte_fallback = spec.get("byte_fallback", False)
+    proto.normalizer_spec.name = "identity"
+    for flag in ("add_dummy_prefix", "remove_extra_whitespaces", "escape_whitespaces"):
+        setattr(proto.normalizer_spec, flag, spec.get(flag, True))
+    piece_type = sp_pb2.ModelProto.SentencePiece.Type
+    for piece, score, typ in spec["pieces"]:
+        proto.pieces.add(piece=piece, score=score, type=piece_type.Value(typ))
+    return proto
+
+
+class Tokenizer:
+    def __init__(self, model):
+        """model: a ModelProto, or tokenizer.json's contents."""
+        proto = model if isinstance(model, sp_pb2.ModelProto) else model_proto(model)
+        serialized = proto.SerializeToString()
+        self.sp = spm.SentencePieceProcessor(model_proto=serialized)
+        # The same model without the dummy prefix or whitespace cleanup, for
+        # tokenizing text fragments byte-for-byte (constrained decoding needs that).
+        fragment = sp_pb2.ModelProto.FromString(serialized)
+        fragment.normalizer_spec.add_dummy_prefix = False
+        fragment.normalizer_spec.remove_extra_whitespaces = False
+        self.sp_fragment = spm.SentencePieceProcessor(model_proto=fragment.SerializeToString())
+
+    @classmethod
+    def from_file(cls, path):
+        """A tokenizer.model, or a tokenizer.json."""
+        path = Path(path)
+        if path.suffix == ".json":
+            return cls(json.loads(path.read_text(encoding="utf-8")))
+        return cls(sp_pb2.ModelProto.FromString(path.read_bytes()))
+
+    def __len__(self):
+        return self.sp.vocab_size()
+
+    def encode(self, text):
+        return self.sp.encode(text)
+
+    def decode(self, ids):
+        return self.sp.decode([int(i) for i in ids])

neudecide-0.0.1.dist-info/METADATA

@@ -0,0 +1,205 @@
+Metadata-Version: 2.4
+Name: neudecide
+Version: 0.0.1
+Summary: NeuDecide - a tiny voice-action model.
+Author-email: neuphonic <general@neuphonic.com>
+License-Expression: Apache-2.0
+Requires-Python: <3.14,>=3.10
+Description-Content-Type: text/markdown
+License-File: LICENSE
+Requires-Dist: huggingface_hub>=0.20
+Requires-Dist: llguidance>=1.8
+Requires-Dist: numpy>=1.24
+Requires-Dist: onnxruntime>=1.23.2
+Requires-Dist: protobuf>=3.20
+Requires-Dist: sentencepiece>=0.2.1
+Requires-Dist: soxr>=0.5
+Dynamic: license-file
+
+# NeuDecide
+
+**A tiny voice-action model. Audio and a list of tools in, a tool call out. No transcript in between.**
+
+[Hugging Face](https://huggingface.co/neuphonic/neudecide) · [PyPI](https://pypi.org/project/neudecide/) · [arxiv (coming soon}]()
+
+NeuDecide converts speech directly into function calls. It's 43 MB, can run on a single thread via ONNX Runtime, and needs no GPU or internet connection once downloaded. Tools are passed in as JSON schemas at inference time, so you change what the model can do by changing the JSON, not by retraining.
+
+```
+"clean the kitchen"  +  [cleanRoom, goToBase, getBattery]   →   [{"name": "cleanRoom", "arguments": {"room": "kitchen"}}]
+```
+
+## Installation
+
+```bash
+pip install neudecide
+```
+
+Requires Python 3.10 – 3.13. The model weights are downloaded from the Hugging Face Hub on first use and cached after that.
+
+## Quickstart
+
+This example turns a recorded voice command into a call for a robot vacuum. Save a WAV recording of yourself saying something like "clean the kitchen" as `command.wav`, then run:
+
+```python
+from neudecide import NeuDecide
+
+TOOLS = [
+    {
+        "name": "cleanRoom",
+        "description": "Vacuum-cleans one room.",
+        "parameters": {
+            "type": "object",
+            "properties": {
+                "room": {
+                    "type": "string",
+                    "enum": [
+                        "living room",
+                        "bathroom",
+                        "kitchen",
+                        "bedroom",
+                    ],
+                },
+            },
+            "required": ["room"],
+        },
+    },
+    {
+        "name": "goToBase",
+        "description": "Sends the vacuum back to its charging base.",
+        "parameters": {
+            "type": "object",
+            "properties": {},
+        },
+    },
+    {
+        "name": "getBattery",
+        "description": "Reports the vacuum's battery level.",
+        "parameters": {
+            "type": "object",
+            "properties": {},
+        },
+    },
+]
+
+
+def main():
+    model = NeuDecide.from_pretrained()
+    calls = model.generate("./examples/command.wav", TOOLS)
+    print(calls)
+
+
+if __name__ == "__main__":
+    main()
+```
+
+Output:
+
+```
+[{'name': 'cleanRoom', 'arguments': {'room': 'kitchen'}}]
+```
+
+`generate` returns a list of `{"name", "arguments"}` dicts, or `[]` when no tool applies.
+
+This script is also in the repo as `examples/basic_example.py`, so from a clone you can run `python examples/basic_example.py`.
+
+## Usage
+
+### Audio input
+
+`audio` can be a path to a PCM WAV file or a NumPy array. Arrays may be mono `(samples,)` or multi-channel `(samples, channels)`; channels are averaged to mono. Audio is resampled to 16 kHz automatically, so pass `sample_rate=` if your array is at a different rate:
+
+```python
+calls = model.generate("command.wav", TOOLS)                     # WAV file
+calls = model.generate(waveform, TOOLS, sample_rate=44100)       # NumPy array
+```
+
+Only WAV is read natively. For other formats (MP3, FLAC, OGG), load the audio yourself and pass the array. Audio longer than 30 seconds is truncated.
+
+### Tool definitions
+
+Each tool has a `name`, a `description`, and `parameters` as a JSON Schema object. You can pass the list directly or as a JSON string. The model only knows what you tell it about each tool, so clear names and descriptions do most of the work. When an argument has a fixed set of values, like the rooms above, list them with `enum`.
+
+## How it works
+
+NeuDecide is a pre-trained streaming English speech encoder joined to a compact tool-calling encoder–decoder, initialised from [Needle](https://github.com/cactus-compute/needle) by Cactus Compute, and trained end to end as one system. Inference runs three ONNX graphs:
+
+| Graph | Runs | Does | Parameters | Size |
+|---|---|---|---|---|
+| `audio_encoder` | Once per utterance | Waveform → 12.5 Hz latent frames | 28.7M | 22.5 MB |
+| `tool_encoder` | Once per utterance and tool list | Audio frames + tool tokens → cross-attention keys and values | 11.8M | 10.1 MB |
+| `decoder_step` | Once per output token | Next token, reusing a cached self-attention state | 15.0M | 10.0 MB |
+
+**TBC** the `tool_encoder` KV-cache state can be retained between tool calls, if the tools remain the same.
+
+| | |
+|---|---|
+| Model width | 512 |
+| Decoder layers | 8 |
+| Attention | Grouped-query, 4 key/value heads of 64 dimensions |
+| Vocabulary | 8,192 SentencePiece BPE tokens |
+| Sample rate | 16 kHz |
+| Limits | 30 s of audio, 1,536 tool tokens, 128 output tokens |
+
+Tools are serialised exactly as Python's `json.dumps` writes them by default, matching how the model was trained.
+
+## Results
+
+<p align="center">
+  <img src="assets/param_bubbles.png" alt="Exact match against tool accuracy on FSC, SNIPS SLU and SLURP for NeuDecide, Whistle → Needle 3, Parakeet 660M → FunctionGemma and Voxtral Mini 3B, with bubble area showing model size" width="720">
+</p>
+
+Exact match against tool accuracy on FSC, SNIPS SLU and SLURP, choosing from 10 tools. Each bubble is one of the four systems in the table below, and its area is the model's size on disk; a cascade counts its ASR and decision model together. ([PDF](assets/param_bubbles.pdf))
+
+We compared NeuDecide with five speech-recognition cascades (Parakeet 110M and 660M feeding Needle or FunctionGemma, and Whistle feeding Needle 3) and with Voxtral Mini 3B, an end-to-end speech model. Each system chooses from 10 tools and must pick the right tool and fill its arguments. The table shows two of the cascades alongside Voxtral Mini 3B; a cascade's size counts its ASR and decision model together.
+
+| | | NeuDecide | Whistle → Needle 3 | Parakeet 660M → FunctionGemma | Voxtral Mini 3B |
+|---|---|---|---|---|---|
+| **Size** | Parameters | 55.5M | 176M | 870M | 3B |
+| **Size** | File size | 43 MB | 52 MB | 3.0 GB | 18.7 GB |
+| **Size** | File size vs NeuDecide | 1× | 1.2× | 71× | 435× |
+| SLURP | Tool accuracy | **77.3%** | 28.5% | 25.7% | 54.4% |
+| SLURP | Exact match | **24.7%** | 3.9% | 1.5% | 7.6% |
+| Fluent Speech Commands | Tool accuracy | **97.0%** | 66.4% | 80.4% | 96.4% |
+| Fluent Speech Commands | Exact match | 82.8% | 58.8% | 65.4% | **95.9%** |
+| SNIPS SLU | Tool accuracy | 70.1% | 61.7% | 43.5% | **87.9%** |
+| SNIPS SLU | Exact match | 20.9% | 15.4% | 17.5% | **36.8%** |
+
+On SLURP, the most varied of the three datasets, NeuDecide beats every other system on every metric. With 10 tools, it gets six times as many commands exactly right as the best cascade and three times as many as Voxtral Mini 3B. On Fluent Speech Commands, it picks the right tool more often than any other system, and it beats every cascade on exact match. Voxtral Mini 3B fills arguments more accurately there, at roughly 50 times NeuDecide's parameter count. SNIPS, with its free-form names, is the weak spot (see [Limitations](#limitations)).
+
+## Performance
+
+Measured with the `q4` export:
+
+| Device | Time to call | Load time |
+|---|---|---|
+| MacBook Pro M3 | 46 ms | 159 ms |
+| Samsung S24+ | 82 ms | 267 ms |
+| Raspberry Pi 5 | 206 ms | 499 ms |
+
+Peak RAM ranged from 146 MB to 174 MB across these devices; it depends on the runtime and workload.
+
+## Limitations
+
+- **It can fire on noise.** Given 30 s of low-level noise, it still produced a call. There is no calibrated confidence score yet, so put a gate in front: voice-activity detection, or an explicit "no action" tool.
+- **Keep tool lists short.** We recommend at most 10 tools. Longer lists also cost more compute than longer audio, and lists over 1,536 tokens are truncated with a warning.
+- **English only.** Inputs are capped at 30 s of audio and 128 output tokens.
+- **Free-form arguments are harder.** On SNIPS, where arguments are often open-ended names such as artists and playlists, transcript-based cascades still fill arguments more accurately. Use `enum` wherever an argument has a fixed set of values.
+
+Test it on your own tools and your users' voices before deploying. NeuDecide works best next to a larger model: let it handle the fast, frequent, bounded decisions, and route anything it can't place to a bigger model or a person.
+
+## Development
+
+```bash
+pip install -e . -r requirements-dev.txt
+pre-commit install
+pytest -m "not slow"   # "slow" tests download the model from the Hub
+```
+
+## Acknowledgements
+
+- The tool-calling encoder–decoder and tokenizer are built on [Needle](https://github.com/cactus-compute/needle) by [Cactus Compute](https://cactuscompute.com/).
+- The frozen ASR model is dervied from [Nvidia FastConformer 32M](https://huggingface.co/nvidia/stt_en_fastconformer_hybrid_medium_streaming_80ms).
+
+## License
+
+Apache 2.0. See [LICENSE](LICENSE).

neudecide-0.0.1.dist-info/RECORD

@@ -0,0 +1,10 @@
+neudecide/__init__.py,sha256=nZICvOHipnynMuyFC9gDHnvdHhWFKHq9NHt2MTwOCcM,76
+neudecide/audio.py,sha256=4MdyqKXgfaEfWks-cAQsIV4GhpeELMRPM4LKaq0Ghc0,2299
+neudecide/grammar.py,sha256=Lg0kkBo-ReCSVlbST183URrEz98ZndU5KVuknMZE8RU,6009
+neudecide/model.py,sha256=K7Vj9jXgMh7Lt8kibRntFAaG4e_kJRtQhGF82TeqXcU,8653
+neudecide/tokenizer.py,sha256=eQOXIlQFhzXAs_4hKpksnmU8KVMtPwGO1B9NRTyj7Gs,2621
+neudecide-0.0.1.dist-info/licenses/LICENSE,sha256=xx0jnfkXJvxRnG63LTGOxlggYnIysveWIZ6H3PNdCrQ,11357
+neudecide-0.0.1.dist-info/METADATA,sha256=PI7LwyCLRKo3zkyvQ2ZTxjT7HM0_cPFTSeWdzBliRwY,9219
+neudecide-0.0.1.dist-info/WHEEL,sha256=YVMoNqKzERt-wjUZwJ33xBGAwnFl-4cqbYkTtWa4itE,91
+neudecide-0.0.1.dist-info/top_level.txt,sha256=3iptrM67QC92r4t3GWrZRNA8vjPlWC8ewGPIkqb6jT4,10
+neudecide-0.0.1.dist-info/RECORD,,

neudecide-0.0.1.dist-info/WHEEL

@@ -0,0 +1,5 @@
+Wheel-Version: 1.0
+Generator: setuptools (84.0.0)
+Root-Is-Purelib: true
+Tag: py3-none-any
+

neudecide-0.0.1.dist-info/licenses/LICENSE

@@ -0,0 +1,201 @@
+                                 Apache License
+                           Version 2.0, January 2004
+                        http://www.apache.org/licenses/
+
+   TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
+
+   1. Definitions.
+
+      "License" shall mean the terms and conditions for use, reproduction,
+      and distribution as defined by Sections 1 through 9 of this document.
+
+      "Licensor" shall mean the copyright owner or entity authorized by
+      the copyright owner that is granting the License.
+
+      "Legal Entity" shall mean the union of the acting entity and all
+      other entities that control, are controlled by, or are under common
+      control with that entity. For the purposes of this definition,
+      "control" means (i) the power, direct or indirect, to cause the
+      direction or management of such entity, whether by contract or
+      otherwise, or (ii) ownership of fifty percent (50%) or more of the
+      outstanding shares, or (iii) beneficial ownership of such entity.
+
+      "You" (or "Your") shall mean an individual or Legal Entity
+      exercising permissions granted by this License.
+
+      "Source" form shall mean the preferred form for making modifications,
+      including but not limited to software source code, documentation
+      source, and configuration files.
+
+      "Object" form shall mean any form resulting from mechanical
+      transformation or translation of a Source form, including but
+      not limited to compiled object code, generated documentation,
+      and conversions to other media types.
+
+      "Work" shall mean the work of authorship, whether in Source or
+      Object form, made available under the License, as indicated by a
+      copyright notice that is included in or attached to the work
+      (an example is provided in the Appendix below).
+
+      "Derivative Works" shall mean any work, whether in Source or Object
+      form, that is based on (or derived from) the Work and for which the
+      editorial revisions, annotations, elaborations, or other modifications
+      represent, as a whole, an original work of authorship. For the purposes
+      of this License, Derivative Works shall not include works that remain
+      separable from, or merely link (or bind by name) to the interfaces of,
+      the Work and Derivative Works thereof.
+
+      "Contribution" shall mean any work of authorship, including
+      the original version of the Work and any modifications or additions
+      to that Work or Derivative Works thereof, that is intentionally
+      submitted to Licensor for inclusion in the Work by the copyright owner
+      or by an individual or Legal Entity authorized to submit on behalf of
+      the copyright owner. For the purposes of this definition, "submitted"
+      means any form of electronic, verbal, or written communication sent
+      to the Licensor or its representatives, including but not limited to
+      communication on electronic mailing lists, source code control systems,
+      and issue tracking systems that are managed by, or on behalf of, the
+      Licensor for the purpose of discussing and improving the Work, but
+      excluding communication that is conspicuously marked or otherwise
+      designated in writing by the copyright owner as "Not a Contribution."
+
+      "Contributor" shall mean Licensor and any individual or Legal Entity
+      on behalf of whom a Contribution has been received by Licensor and
+      subsequently incorporated within the Work.
+
+   2. Grant of Copyright License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      copyright license to reproduce, prepare Derivative Works of,
+      publicly display, publicly perform, sublicense, and distribute the
+      Work and such Derivative Works in Source or Object form.
+
+   3. Grant of Patent License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      (except as stated in this section) patent license to make, have made,
+      use, offer to sell, sell, import, and otherwise transfer the Work,
+      where such license applies only to those patent claims licensable
+      by such Contributor that are necessarily infringed by their
+      Contribution(s) alone or by combination of their Contribution(s)
+      with the Work to which such Contribution(s) was submitted. If You
+      institute patent litigation against any entity (including a
+      cross-claim or counterclaim in a lawsuit) alleging that the Work
+      or a Contribution incorporated within the Work constitutes direct
+      or contributory patent infringement, then any patent licenses
+      granted to You under this License for that Work shall terminate
+      as of the date such litigation is filed.
+
+   4. Redistribution. You may reproduce and distribute copies of the
+      Work or Derivative Works thereof in any medium, with or without
+      modifications, and in Source or Object form, provided that You
+      meet the following conditions:
+
+      (a) You must give any other recipients of the Work or
+          Derivative Works a copy of this License; and
+
+      (b) You must cause any modified files to carry prominent notices
+          stating that You changed the files; and
+
+      (c) You must retain, in the Source form of any Derivative Works
+          that You distribute, all copyright, patent, trademark, and
+          attribution notices from the Source form of the Work,
+          excluding those notices that do not pertain to any part of
+          the Derivative Works; and
+
+      (d) If the Work includes a "NOTICE" text file as part of its
+          distribution, then any Derivative Works that You distribute must
+          include a readable copy of the attribution notices contained
+          within such NOTICE file, excluding those notices that do not
+          pertain to any part of the Derivative Works, in at least one
+          of the following places: within a NOTICE text file distributed
+          as part of the Derivative Works; within the Source form or
+          documentation, if provided along with the Derivative Works; or,
+          within a display generated by the Derivative Works, if and
+          wherever such third-party notices normally appear. The contents
+          of the NOTICE file are for informational purposes only and
+          do not modify the License. You may add Your own attribution
+          notices within Derivative Works that You distribute, alongside
+          or as an addendum to the NOTICE text from the Work, provided
+          that such additional attribution notices cannot be construed
+          as modifying the License.
+
+      You may add Your own copyright statement to Your modifications and
+      may provide additional or different license terms and conditions
+      for use, reproduction, or distribution of Your modifications, or
+      for any such Derivative Works as a whole, provided Your use,
+      reproduction, and distribution of the Work otherwise complies with
+      the conditions stated in this License.
+
+   5. Submission of Contributions. Unless You explicitly state otherwise,
+      any Contribution intentionally submitted for inclusion in the Work
+      by You to the Licensor shall be under the terms and conditions of
+      this License, without any additional terms or conditions.
+      Notwithstanding the above, nothing herein shall supersede or modify
+      the terms of any separate license agreement you may have executed
+      with Licensor regarding such Contributions.
+
+   6. Trademarks. This License does not grant permission to use the trade
+      names, trademarks, service marks, or product names of the Licensor,
+      except as required for reasonable and customary use in describing the
+      origin of the Work and reproducing the content of the NOTICE file.
+
+   7. Disclaimer of Warranty. Unless required by applicable law or
+      agreed to in writing, Licensor provides the Work (and each
+      Contributor provides its Contributions) on an "AS IS" BASIS,
+      WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
+      implied, including, without limitation, any warranties or conditions
+      of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
+      PARTICULAR PURPOSE. You are solely responsible for determining the
+      appropriateness of using or redistributing the Work and assume any
+      risks associated with Your exercise of permissions under this License.
+
+   8. Limitation of Liability. In no event and under no legal theory,
+      whether in tort (including negligence), contract, or otherwise,
+      unless required by applicable law (such as deliberate and grossly
+      negligent acts) or agreed to in writing, shall any Contributor be
+      liable to You for damages, including any direct, indirect, special,
+      incidental, or consequential damages of any character arising as a
+      result of this License or out of the use or inability to use the
+      Work (including but not limited to damages for loss of goodwill,
+      work stoppage, computer failure or malfunction, or any and all
+      other commercial damages or losses), even if such Contributor
+      has been advised of the possibility of such damages.
+
+   9. Accepting Warranty or Additional Liability. While redistributing
+      the Work or Derivative Works thereof, You may choose to offer,
+      and charge a fee for, acceptance of support, warranty, indemnity,
+      or other liability obligations and/or rights consistent with this
+      License. However, in accepting such obligations, You may act only
+      on Your own behalf and on Your sole responsibility, not on behalf
+      of any other Contributor, and only if You agree to indemnify,
+      defend, and hold each Contributor harmless for any liability
+      incurred by, or claims asserted against, such Contributor by reason
+      of your accepting any such warranty or additional liability.
+
+   END OF TERMS AND CONDITIONS
+
+   APPENDIX: How to apply the Apache License to your work.
+
+      To apply the Apache License to your work, attach the following
+      boilerplate notice, with the fields enclosed by brackets "[]"
+      replaced with your own identifying information. (Don't include
+      the brackets!)  The text should be enclosed in the appropriate
+      comment syntax for the file format. We also recommend that a
+      file or class name and description of purpose be included on the
+      same "printed page" as the copyright notice for easier
+      identification within third-party archives.
+
+   Copyright [yyyy] [name of copyright owner]
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+       http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License.

neudecide-0.0.1.dist-info/top_level.txt

@@ -0,0 +1 @@
+neudecide