Voice to Data: LLM Pipeline for Lab Notebooks

Python
LLM
NLP
MLOps
Lab Automation
A voice-driven pipeline that turns laboratory bench recordings into structured, validated records — ASR, LLM extraction, domain validation, and a GxP-compliant audit trail. Built for pharma and biotech R&D.
Published

February 22, 2026

Overview

Scientists in formulation, analytical, and cell culture labs spend 20–40 minutes per bench session transcribing observations into ELN/LIMS systems — after the experiment, from memory, under time pressure. This pipeline removes the keyboard from the critical path.

The pipeline takes an audio file recorded at the bench and produces a structured, validated JSON record ready for ELN/LIMS handoff, with an immutable audit trail that satisfies 21 CFR Part 11 requirements.

All processing runs locally — no audio or text leaves your network, which is the only defensible configuration for regulated R&D environments.

Architecture

Voice (audio file)
     │
     ▼
┌──────────────────────────────┐
│  Stage 1 — ASR               │
│  faster-whisper (local)      │
└──────────────┬───────────────┘
               │
               ▼
┌──────────────────────────────┐
│  Stage 2 — Extraction        │
│  Instructor + Pydantic       │
│  + Ollama (local LLM)        │
└──────────────┬───────────────┘
               │
               ▼
┌──────────────────────────────┐
│  Stage 3 — Validation        │
│  DomainValidator             │
│  (plausibility + mandatory)  │
└──────────────┬───────────────┘
               │
               ▼
┌──────────────────────────────┐
│  Stage 4 — Serialization     │
│  JSON + AuditRecord          │
└──────────────────────────────┘

Stack

Layer Tool Rationale
ASR faster-whisper (CTranslate2) Open-weights, 2–4× faster than PyTorch Whisper, CPU-capable
LLM structured extraction instructor + ollama Guaranteed schema-conformant output via tool-calling — no JSON parsing brittleness
Schema & validation pydantic v2 Type enforcement + domain plausibility checks in one place
Dependency management uv Fast, reproducible, lockfile-based

Key design decisions

Why Instructor over raw LLM JSON? Asking an LLM to “return JSON” produces unreliable free-form output. Instructor uses the model’s tool-calling mode to force every field of the Pydantic schema to be explicitly populated. Missing fields become null rather than hallucinated values — which matters a great deal in a compliance context.

Two-tier validation: Pydantic enforces structural correctness (types, enums). A separate DomainValidator class enforces scientific plausibility (pH ∈ [0, 14], viscosity ranges, unit whitelists). These are kept separate intentionally: structural failures are hard blocks; domain flags are soft and route the record to human review.

Hedged values as a first-class concept: When a scientist says “approximately 4,200 millipascal seconds”, the hedged=True flag is set in the schema and the record goes to the review queue — not committed blindly. This is a deliberate design choice, not a limitation.

Compliance notes

  • Audit trail: every record carries a SHA-256 hash of the raw transcript, model version strings, extraction timestamp, and review status — nothing needed for a 21 CFR Part 11 audit is missing
  • Model version-locking: models are pinned; updates are treated as change control events requiring requalification
  • Temperature = 0: extraction calls are as deterministic as the model allows
  • Local-only: no cloud API calls, no data processing agreement required
Back to top