Ejecuta cualquier Skill en Manus
con un clic

Ejecuta cualquier Skill en Manus con un clic

Comenzar

cohere-stt

Estrellas0

Forks0

Actualizado12 de julio de 2026 a las 01:54

Speech-to-Text using Cohere Transcribe (2B, Apache 2.0)

Instalación

Instalar con Codex o Claude Copia este prompt, pégalo en Codex, Claude u otro asistente, y deja que revise la página de la skill y la instale por ti.

Ejecutar en Manus

Fuente

2lab-ai

2lab-ai/skills

Abrir repositorio de GitHub Ver repositorios del creador

Descarga

Ejecutar en Manus

Ocupaciones relacionadasSOC

Basado en la clasificación ocupacional SOC

Desarrolladores de softwareOcupaciones informáticas y matemáticas·SOC 15-1252

Explorador de archivos

23 archivos

SKILL.md

readonly

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

No auto language detection — must specify --language
Single language per call — no code-switching support
Eager transcription — transcribes non-speech sounds (use VAD pre-filter for noisy audio)
No timestamps — output is plain text only (no word-level timing)
No speaker diarization — single-speaker transcription
First --compile call is slow — subsequent calls benefit from torch.compile cache
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

Más de este repositorio

mismo repositorio

beautiful-pdf

2lab-ai/skills

Convert markdown to a beautifully designed editorial-style PDF report. Uses Pretendard variable font, Swiss-Coffee + Cool-Blue palette, minimalist-maximalism layout, cover page, table of contents, section dividers, page numbers. Optimized for Korean+English mixed reports. Use when user asks for "report PDF", "보고서 PDF", "아름다운 PDF", "예쁜 PDF", or wants to publish a markdown document as a polished document.

2026-07-120

blog-shortform

2lab-ai/skills

블로그 URL 1개를 받아 60초 9:16 숏폼 영상(mp4)을 자동 생성하는 파이프라인. 한국 블로그 4종(티스토리/네이버/벨로그/브런치)에서 본문·이미지를 추출하고, 2인 대화 대본을 만든 뒤 로컬 fish-tts(soma-voice HTTP)와 GPT Image 2(gpt-image-2)로 음성·이미지를 생성해 Remotion v4로 합성한다. "블로그 URL로 숏폼 만들어줘", "이 글로 릴스 영상 만들어줘", "blog to shortform", "blog-shortform", "blog-url-to-shortform" 같은 요청에서 발동한다.

2026-07-120

create-cli

2lab-ai/skills

CLI UX/spec: args, flags, help, output, errors, config, dry-run.

2026-07-120

dialogue-archive

2lab-ai/skills

Use when 지혁 asks to record/save/archive a conversation between him and Elon — produces (1) verbatim RAW markdown, (2) cleaned-up EDITED markdown with nothing removed, and (3) a messenger-style HTML blog post. Triggers on "대화 기록해줘", "이거 저장해서 보내줘", "블로그에 올리게 html로", "원문 그대로 + 교정본 + 포스트".

2026-07-120

frontend-design

2lab-ai/skills

Production frontend UI: pages, apps, components, polished non-generic design.

2026-07-120

github-deep-review

2lab-ai/skills

GitHub deep review: bugs, PRs, best fix, stale-or-real, read-code-first.

2026-07-120

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

No auto language detection — must specify --language
Single language per call — no code-switching support
Eager transcription — transcribes non-speech sounds (use VAD pre-filter for noisy audio)
No timestamps — output is plain text only (no word-level timing)
No speaker diarization — single-speaker transcription
First --compile call is slow — subsequent calls benefit from torch.compile cache
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