RobotsMali/soloni-ic-slot-fintech-v0

Soloni IC/Slot Fintech v0

soloni-ic-slot-fintech-v0 is an end-to-end Bambara spoken-language-understanding model for a small mobile-banking ontology. It maps audio directly to a structured representation with scenario, action, and entities, without requiring a separate transcript or general-purpose language model.

The checkpoint powers the SLU path in `mobilebamspeech` and the `mobileBAMking` research/UX demonstration. Its multi-graph Android inference implementation is available in `NeMoOnnxSharp`.

Supported Domain

The collection contains these actions:

• —transfer_to_account and transfer_to_momo, with amount and account/phone-number entities;
• —get_balance;
• —Maps_to, with a destination-page entity;
• —logout;
• —search_faq, with topic/category entities.

Consumers must treat this as a closed ontology and reject unsupported or malformed output. It is not a conversational banking agent.

Use with NVIDIA NeMo

pip install "nemo-toolkit[asr]"

from nemo.collections.asr.models import SLUIntentSlotBPEModel

model = SLUIntentSlotBPEModel.from_pretrained(
    "RobotsMali/soloni-ic-slot-fintech-v0"
)
# SLUIntentSlotBPEModel needs these compatibility patches in NeMo 2.x.
model.decoder.freeze = lambda: None
model.decoder.unfreeze = lambda *args, **kwargs: None

original_decode = model.sequence_generator.decode_semantics_from_tokens

def patched_decode(seq_tokens):
    # NeMo 2.x may return (tokens, lengths); decode the token tensor.
    if isinstance(seq_tokens, tuple):
        seq_tokens = seq_tokens[0]
    return original_decode(seq_tokens)

model.sequence_generator.decode_semantics_from_tokens = patched_decode

model.eval()
predictions = model.transcribe(["banking_command.wav"])
print(predictions[0])

Apply both patches after loading the model and before the first call to transcribe(). They work around NeMo 2.x incompatibilities in decoder freezing and tuple-valued sequence-generator output; they were tested with NeMo 2.5.0. The model preprocesses speech at 16 kHz. Its decoder emits a Python-style serialized structure such as a mapping containing scenario, action, and an entities list. Parse defensively: validate syntax, allow-listed actions, entity types, and application state before executing anything.

Architecture

The model uses NeMo's SLUIntentSlotBPEModel:

• —a 17-layer Conformer encoder (d_model=512) initialized from `soloni-114m-tdt-ctc-v3`;
• —a 58-piece SentencePiece tokenizer;
• —a three-layer, eight-head Transformer decoder;
• —a token classifier trained with negative log-likelihood.

The .nemo archive is approximately 475 MB. The architecture is based on NVIDIA NeMo's speech intent and slot-filling example.

Training Data and Procedure

The archived collection contains 588 annotated Bambara voice commands: 361 Operation, 222 Navigate, and 5 FAQ examples. It is strongly imbalanced and was collected for the MobileBAMking prototype. The preparation pipeline and raw export are in `slu-ic-slot-filling/`; referenced recordings are downloaded when manifests are built.

Training used NeMo 2.5.0, batches of 16, an unfrozen pretrained encoder, AdamW, a 5e-4 decoder learning rate, a 1e-4 encoder learning rate, cosine annealing, and up to 50 epochs. See the released config.

Evaluation

The repository retains the complete text output from a 60-example held-out evaluation:

All 60 predictions were syntactically valid. These results are micro-averaged and in-domain; the tiny FAQ class and small test set do not establish broad coverage.

Limitations and Safety

• —The model recognizes only the application ontology and may confidently misroute unfamiliar speech.
• —Training data are small, imbalanced, and tied to one collection workflow.
• —Entity recall is materially lower than intent accuracy; never assume a missing or malformed amount/account value.
• —This checkpoint does not authenticate a user, verify a speaker, confirm consent, or authorize financial activity.
• —Any real application must use explicit confirmation, deterministic validation, secure backend controls, audit logging, and broader evaluation.

License

The model is released under CC BY 4.0. The source voice collection may have additional privacy or distribution constraints and is not published as a standalone dataset.