title	TranscriptAI
emoji	🧠
colorFrom	pink
colorTo	red
sdk	docker
app_port	7860
pinned	false

Multilingual Meeting Intelligence · Japanese · Hindi · English · Mixed

Turns any meeting transcript into structured business intelligence in ~3 seconds.

Turns any meeting transcript into structured business intelligence in ~3 seconds.

The Problem No Generic AI Tool Solves

Generic meeting summarisers extract what was said. They miss what was meant.

What was said	Generic AI	TranscriptAI
検討いたします	"Action: We will consider it"	⚠ Soft rejection — 72% confidence. Follow up explicitly.
難しいかもしれません	"It may be difficult" — neutral	🚨 HIGH rejection signal — 90% confidence. Deal at risk.
前向きに検討	"Positive consideration"	⚠ Uncertain — outcome not guaranteed (55%)
承知いたしました	"Acknowledged"	🏯 High keigo register — senior speaker or formal context
善処します	"Action: We will handle it"	🚨 Classic nemawashi dodge — no real commitment made
パートナーシップは継続しないことを決定しました	"Decision made"	⛔ CRITICAL — Explicit contract termination. Irrevocable.

Japanese enterprise also mandates APPI compliance — raw meeting data cannot be sent to foreign cloud LLMs. Most tools fail this requirement by design. TranscriptAI masks PII locally before any LLM call.

---

What It Does

Input: Any transcript (JP · HI · EN · Mixed) or Audio file (MP3/MP4/WAV)
Output: Structured business intelligence in ~3 seconds

• Trilingual analysis — Japanese / Hindi (Hinglish) / English / Mixed in a single meeting
• Keigo detection — MeCab morphological analysis extracts formality register (High / Medium / Low)
• 20 soft rejection patterns — nemawashi, 難しいですね, ぜひ検討, and more with confidence scores
• CRITICAL risk tier — explicit contract termination detection, separate from soft refusals
• 8 Hindi indirect patterns — देखते हैं, थोड़ा मुश्किल है, कोशिश करेंगे and more
• APPI-compliant PII masking — 500+ Japanese surnames, local NER, before any LLM call
• Hallucination guard — 100% rule-based token overlap; LLM never validates its own output
• Meeting Health Score — 0–100 across sentiment, action clarity, communication risk, AI confidence
• 議事録 export — structured Japanese business minutes in standard enterprise format
• Cultural insights export — 根回し risk, 稟議 approval status, nemawashi pattern breakdown
• Groq Whisper — MP4 / MP3 / WAV audio transcription
• Vector cache — ChromaDB semantic similarity for instant return on similar transcripts
• 2-key Groq rotation — auto-failover on 429 rate limits

What was said	Generic AI	TranscriptAI
検討いたします	"Action: We will consider it"	⚠ Soft rejection — 72% confidence
難しいかもしれません	"It may be difficult" — neutral	🚨 HIGH rejection signal — deal at risk
パートナーシップは継続しないことを決定しました	"Decision made"	⛔ CRITICAL — Explicit contract termination

Accuracy History

Version	Key change	Accuracy
v1	Hard exact matching, English-only	22–30%
v2	Fuzzy speaker names, TF-IDF similarity	~45%
v3	MeCab keigo override, bilingual ground truth	~60%
v4	Hallucination guard, nemawashi patterns, APPI masking	75–85%
v5 (live)	2-key rotation, vector cache, MLflow, bypass_cache eval fix	93.8%

Every rebuild was driven by evaluation metric failures traced through the pipeline — not intuition. When action F1 was 0.4 at v2, tracing showed the owner field extracted role titles ("Director") instead of first names. One prompt instruction fixed it.

---

Evaluation Metrics · v5 Live

Test case	Overall	ROUGE-1	Action F1	Sentiment
Sales call · JA/EN mixed	94.5%	0.694	1.0	1.0
Internal meeting · Japanese	93.8%	0.703	1.0	1.0
Client conflict · EN/JA	93.8%	0.703	1.0	1.0

---

Architecture — 11-Stage Pipeline

Input transcript
    │
    ▼
 1  Vector cache check       utils/vector_cache.py          ChromaDB cosine similarity
 2  MD5 exact cache          utils/cache.py                 Instant return if identical
 3  PII masking              transcription/pii_masker.py    APPI — BEFORE LLM
 4  LLM analysis             analysis/analyzer.py           Groq key1 → key2 → mock
 5  PII restoration          transcription/pii_masker.py    BEFORE speaker normalization
 6  Speaker normalization    transcription/speaker_normalizer.py
 7  MeCab keigo override     analysis/japanese_tokenizer.py
 8  Code-switch count        utils/evaluator.py             Rule-based Unicode ranges
 9  Hallucination guard      analysis/hallucination_guard.py
10  Soft rejection detection analysis/soft_rejection_detector.py
11  Cache + log              utils/vector_cache.py + utils/logger.py

Critical ordering constraint: PII must be masked before the LLM sees the text. PII must be restored before speaker normalization — otherwise the normalizer receives [NAME_1] and cannot resolve 田中 ↔ Tanaka ↔ Director as the same person.

---

Technology Decisions

Decision	Chosen	Over	Reason
Vector DB	ChromaDB	Pinecone, Weaviate	Free, local, HF Spaces compatible, APPI compliant
LLM inference	Groq	OpenAI, Anthropic	10–20× faster (LPU vs GPU). Free tier. JSON mode.
Japanese NLP	MeCab + IPADIC	spaCy ja, Fugashi	Morpheme-level auxiliary verb detection — keigo is invisible to word-level tokenizers
Web framework	FastAPI	Flask, Django	Native async with asyncio.to_thread(). Auto Swagger.
Eval tracking	MLflow	W&B, Neptune	Free, local SQLite, APPI compliant

---

Quick Start

git clone https://github.com/aiKunalBisht/Transcript-ai.git
cd Transcript-ai
pip install -r requirements.txt
export GROQ_API_KEY=your_key_here

# FastAPI (docs at localhost:7860/docs)
uvicorn main:app --reload --port 7860

Local inference (zero cost, full APPI data residency):

ollama pull qwen3:8b
# App auto-detects Ollama — no config needed

---

Accuracy

main.py                           FastAPI server + route handlers
analysis/
  analyzer.py                     LLM orchestration + 2-key Groq rotation
  soft_rejection_detector.py      20 JP nemawashi + 8 HI indirect + CRITICAL termination
  japanese_tokenizer.py           MeCab keigo extraction
  hallucination_guard.py          Rule-based token overlap validation
  english_analyzer.py             English NLP patterns
  hindi_analyzer.py               Hindi/Hinglish indirect signals
agents/
  gijiroku_formatter.py           議事録 Japanese business minutes generator
  cultural_insights_formatter.py  Cultural context layer for JP meetings
  slide_architect.py              PPTX slide plan generation
exporters/
  pptx_builder.py                 PowerPoint export builder
transcription/
  pii_masker.py                   APPI compliance — local PII masking
  audio_processor.py              Groq Whisper audio transcription
utils/
  html_renderer.py                Results renderer — health score, tabs, insights
  vector_cache.py                 ChromaDB semantic cache
  logger.py                       JSONL trends + drift detection
templates/
  base.html                       Sidebar, navigation, session persistence
  index.html                      Main analysis page
  export.html                     Export page — PPTX, 議事録, MD, JSON, TXT
static/
  style.css                       Full responsive CSS — mobile, tablet, desktop
tests/
  test_data.py                    3 bilingual ground truth test cases

---

REST API

# Analyze a transcript
POST /analyze-text
  transcript: str
  language:   str | null   # auto-detect if null
  mask_pii:   bool         # default true

# Exports
POST /export/pptx
POST /export/gijiroku
POST /export/cultural-insights
POST /export/markdown
POST /export/json
POST /export/txt

# Health
GET /health

---

Performance

Metric	Score
Lighthouse Performance	94
Lighthouse Accessibility	100
CLS (Cumulative Layout Shift)	0
Speed Index	1.2s

---

Built by Kunal Bisht AI Engineer · LLM Systems & RAG Pipelines · Multilingual NLP Uttarakhand, India · Open to Remote / Relocation