cohere-stt

name	cohere-stt
description	Speech-to-Text using Cohere Transcribe (2B, Apache 2.0)
triggers	["음성 인식","STT","음성을 텍스트로","transcribe","speech to text"]

Cohere STT - Speech-to-Text Skill

음성 파일을 텍스트로 변환하는 스킬. Cohere Transcribe (2B params) 모델 사용. Open ASR Leaderboard 1위, 한국어 포함 14개 언어 지원.

Quick Start

Single File

bash ~/2lab.ai/skills/cohere-stt/scripts/cohere-stt.sh /path/to/audio.wav --language ko

Multiple Files (Batch)

bash ~/2lab.ai/skills/cohere-stt/scripts/cohere-stt.sh file1.wav file2.mp3 file3.ogg --language ko --json

JSON Output

bash ~/2lab.ai/skills/cohere-stt/scripts/cohere-stt.sh audio.wav --json --language ko

Architecture

Audio File (WAV/MP3/OGG/FLAC)
  → Cohere Transcribe (2B Conformer + Transformer Decoder)
  → Text Output

Pipeline:
1. Audio → 16kHz mono resample (automatic)
2. Mel-spectrogram extraction
3. Conformer encoder (majority of 2B params)
4. Lightweight Transformer decoder → text tokens

Parameters

Flag	Default	Description
--language, -l	ko	ISO 639-1 code: en,de,fr,it,es,pt,el,nl,pl,ar,vi,zh,ja,ko
--device, -d	auto	auto, cuda:0, cpu
--batch-size, -b	16	GPU batch size
--compile	off	torch.compile for faster throughput (warmup cost)
--no-punctuation	off	Remove punctuation from output
--json, -j	off	Output as JSON with stats
--output, -o	stdout	Save to file

Supported Languages (14)

Code	Language	Code	Language
en	English	el	Greek
de	German	nl	Dutch
fr	French	pl	Polish
it	Italian	ar	Arabic
es	Spanish	vi	Vietnamese
pt	Portuguese	zh	Chinese
ja	Japanese	ko	Korean

API Server (for soma integration)

# Start API server (OpenAI Whisper-compatible endpoint)
~/.cohere-stt-venv/bin/python ~/2lab.ai/skills/cohere-stt/scripts/transcribe_api.py --port 8787

# Test
curl -X POST http://localhost:8787/v1/audio/transcriptions \
  -F "file=@audio.ogg" \
  -F "language=ko"

Benchmark

# Run full benchmark (generate 100 TTS samples + transcribe + evaluate)
~/.cohere-stt-venv/bin/python ~/2lab.ai/skills/cohere-stt/scripts/benchmark.py --all

# Individual steps
~/.cohere-stt-venv/bin/python ~/2lab.ai/skills/cohere-stt/scripts/benchmark.py --generate
~/.cohere-stt-venv/bin/python ~/2lab.ai/skills/cohere-stt/scripts/benchmark.py --transcribe
~/.cohere-stt-venv/bin/python ~/2lab.ai/skills/cohere-stt/scripts/benchmark.py --evaluate

Performance

• WER: 5.42 avg on Open ASR Leaderboard (English)
• Speed: RTFx up to 3x (3 seconds of audio processed in 1 second)
• Model Load: ~10-15s on RTX 4090
• Memory: ~6-8GB VRAM (2B params, fp32)
• Long audio: Auto-chunked at 35s segments with overlap

Constraints & Gotchas

1. No auto language detection — must specify --language
2. Single language per call — no code-switching support
3. Eager transcription — transcribes non-speech sounds (use VAD pre-filter for noisy audio)
4. No timestamps — output is plain text only (no word-level timing)
5. No speaker diarization — single-speaker transcription
6. First --compile call is slow — subsequent calls benefit from torch.compile cache
7. GPU memory — ~6-8GB VRAM required; reduce batch_size if OOM

File Structure

cohere-stt/
├── SKILL.md                    # This file
├── scripts/
│   ├── cohere-stt.sh           # Shell wrapper (entry point)
│   ├── inference.py            # Main inference script
│   ├── transcribe_api.py       # HTTP API server (OpenAI-compatible)
│   └── benchmark.py            # TTS→STT benchmark tool
├── examples/                   # Example audio files
└── output/                     # Benchmark results
    ├── benchmark/              # Generated TTS samples
    ├── benchmark_samples.json  # Sample metadata
    └── benchmark_results.json  # Evaluation results

Dependencies

• Python venv: ~/.cohere-stt-venv/
• Model: CohereLabs/cohere-transcribe-03-2026 (auto-downloaded from HuggingFace)
• Packages: transformers>=4.56, torch, huggingface_hub, soundfile, librosa, sentencepiece, protobuf
• GPU: NVIDIA GPU with CUDA (RTX 4090 Laptop tested)
• License: Apache 2.0

Model Card

• Publisher: Cohere Labs
• Release: March 26, 2026
• Parameters: 2B
• Architecture: Conformer encoder + Transformer decoder
• Training Data: 500K hours curated audio-transcript pairs
• HuggingFace: https://huggingface.co/CohereLabs/cohere-transcribe-03-2026