Cohere Local PoC — Audio Transcription + LLM Cleanup + Summary

Language / 语言： English | 中文

A self-contained local feasibility workspace for running CohereLabs/cohere-transcribe-03-2026 on Apple Silicon (tested on Mac mini M4 16 GB) and piping the output through an optional LLM-based cleanup plus one structured post-processing output: an analysis pack.

The workspace answers one concrete question:

Can a Mac run the Cohere model locally well enough to justify integrating a summary-only-fast pipeline later?

---

Pipeline overview

input/audio.wav
    │
    ▼
poc_cohere_local_transcribe.py   ← local ASR (transformers + MPS)
    │  output/transcript.txt
    ▼
transcript_cleanup.py            ← optional LLM dedup / punctuation pass
    │  output/transcript_cleaned.txt
    ▼
transcript_summary.py            ← LLM analysis pack (can read transcript.txt or transcript_cleaned.txt)
       output/transcript_cleaned_analysis.md

Each stage is an independent script and can be run in isolation.

---

Layout

.
├── .env                   ← local secrets (git-ignored)
├── .env.example           ← copy this, fill in your keys
├── .venv/                 ← created by setup_venv.sh (git-ignored)
├── input/                 ← put your audio files here (git-ignored)
├── logs/                  ← runtime logs (git-ignored)
├── models/                ← downloaded model weights (git-ignored)
│   └── CohereLabs/
├── output/                ← generated transcripts and summaries (git-ignored)
├── requirements.txt
├── scripts/
│   ├── autoFull.py
│   ├── autoFull.sh
│   ├── download_model.sh
│   ├── setup_venv.sh
│   ├── poc_cohere_local_transcribe.py
│   ├── transcript_cleanup.py
│   └── transcript_summary.py
├── README.md
└── README.zh.md

---

Requirements

• Python 3.11 — the local ASR stack is more compatible on 3.11 than on newer versions.
• An LLM API key for the cleanup and summary steps (see Local env).

---

Setup

./scripts/setup_venv.sh

The script auto-detects python3.11 (or falls back to python3). You can override the interpreter:

PYTHON=/opt/homebrew/bin/python3.11 ./scripts/setup_venv.sh

---

Model download

./scripts/download_model.sh

The model is saved to ./models/CohereLabs/cohere-transcribe-03-2026. The PoC script detects this path automatically.

To use a mirror (e.g. inside mainland China):

HF_ENDPOINT=https://hf-mirror.com ./scripts/download_model.sh

---

Local env

Copy the template and fill in your API key:

cp .env.example .env

.env.example:

LLM_API_KEY=
LLM_BASE_URL=https://api.siliconflow.cn/v1
LLM_MODEL=deepseek-ai/DeepSeek-V3

LLM_BASE_URL and LLM_MODEL are optional — the defaults above are used when they are absent.

---

Run

Full pipeline (transcribe → optional cleanup → summary)

The quickest way to run everything in one shot is the Python CLI:

./scripts/autoFull.py

The shell wrapper is kept for compatibility:

./scripts/autoFull.sh

Cleanup is skipped by default. To enable step 2, pass:

./scripts/autoFull.sh --enCleanUp 1

All options are optional — defaults are picked up automatically:

./scripts/autoFull.sh \
  --input    ./input/extracted_audio.wav \
  --output   ./output \
  --language en \
  --model    ./models/CohereLabs/cohere-transcribe-03-2026 \
  --env      ./.env

-h / --help prints the full usage.

Step by step

If you need finer control, run each stage individually.

Transcribe only

./.venv/bin/python ./scripts/poc_cohere_local_transcribe.py \
  --input ./input/extracted_audio.wav \
  --output-dir ./output

Cleanup only

./.venv/bin/python ./scripts/transcript_cleanup.py \
  --input ./output/transcript.txt \
  --output ./output/transcript_cleaned.txt

Summary only

./.venv/bin/python ./scripts/transcript_summary.py \
  --input ./output/transcript_cleaned.txt

Optional: explicitly set the analysis output path:

./.venv/bin/python ./scripts/transcript_summary.py \
  --input ./output/transcript_cleaned.txt \
  --output ./output/transcript_cleaned_analysis.md

---

Output files

File	Description
output/transcript.txt	Raw ASR output
output/transcript_cleaned.txt	LLM-deduped and punctuated transcript
output/transcript_analysis.md	Analysis pack generated directly from the raw transcript
output/transcript_cleaned_analysis.md	Claim/evidence/risk analysis pack + X thread draft
output/report.json	Runtime benchmark — model path, Python/torch version, device, audio duration, load time, transcription time, RSS
logs/poc_cohere_local_transcribe.log	ASR run log
logs/transcript_cleanup.log	Cleanup run log
logs/transcript_summary.log	Summary run log

---

Notes

• If you see multilingual garbage in the transcription output, the most likely cause is an unsupported transformers version. Check requirements.txt for the tested version range.
• The LLM cleanup and summary steps are independent modules — they can be imported or called without the ASR step.
• The summary stage now produces one Markdown file: an analysis pack for claim / evidence / risk review.
• Final Markdown is post-processed for readability, with sentence-first line breaks after . ? ! and 。？！ where possible.