Good evening, Kappa-dappa!(o’Θ’o)ﾉ))

Um… I said, “It’s Kappa!”ヾ(･Θ･` )

Anyway, listen—in games like Survival and PvP games, I keep running into Earthlings who say stuff I don’t understand!

I’m not good at either Japanese or Earth languages, so it’s always a hassle! ┐(゜Θ。)┌ Oh well…

So, I recently set things up so I can translate Earth languages other than Japanese into Japanese—without using my home computer—and display them on screen via OBS or text!

I was so happy, thinking, “Now I can watch foreign shows without subtitles with peace of mind!”

But…

My original goal was to have it recognize the words I spoke as Japanese and translate them into other languages…

While it’s 80–90% accurate with English and standard Japanese speech, it stops working completely when I use it on my own voice!

Darn it! It’s Earth’s AI and speech recognition! Kappa wasn’t included in the system after all! XDD

So, that left me feeling pretty down.

Well, never mind that.

You guys still play games and browse all kinds of media, right?

You don’t play for 8 hours a day, do you?

You don’t use up all the free tokens on that service called Cloudflare, do you?

You’re not such a chatterbox that you talk nonstop for 6 hours a day, are you?

OK!

In that case, why not try using this to decipher random comments from people outside the English-speaking world?

Hopefully, you don’t live in the Tower of Babel.

This is Kappa!

Below is an explanation from the AI you’ve all been trusting lately!

Just so you know, I’m not responsible if you use this—even if it blows up or humanity goes extinct, it’s not my fault.

---

OBS + Cloudflare Workers AI + AI Gateway Subtitle Script v6.3 (0805)

Real-time subtitle translation script for OBS streaming. Combines Cloudflare Workers AI (Whisper speech recognition + LLM translation) with AI Gateway for low-latency, low-cost subtitle generation. Runs entirely on the Cloudflare Workers AI free tier (10,000 Neurons/day).

---

Table of Contents

1. Features
2. Requirements
3. Cloudflare Account Setup
4. Running the Script
5. OBS Configuration
6. Use Cases
7. Whisper Model Selection
8. Troubleshooting
9. Version History
10. License
11. Disclaimer
12. Differences from the Japanese Version

---

Features

Speech Recognition & Translation

Feature	Description
Whisper speech recognition	Uses @cf/openai/whisper (multilingual), @cf/openai/whisper-tiny-en (English-only, fast), or @cf/openai/whisper-large-v3-turbo (high-accuracy) on Cloudflare Workers AI
LLM translation	Translates via @cf/meta/llama-3.1-8b-instruct-fast. Supports Japanese to English, English to Japanese, bidirectional auto-detect
AI Gateway routing	All API requests go through AI Gateway for caching, logging, and rate limiting
Prompt Caching	x-session-affinity header caches common prefix for reduced TTFT + Neuron savings (v6.2+)
Token usage display	[Token] in=X out=Y shows real-time Neuron consumption (v6.2+)

Noise Removal & VAD

Feature	Description
Dynamic noise gate	Learns ambient noise for 30 seconds at startup, auto-sets noise floor (mic mode only)
Frequency VAD	Judges human voice by 200-3400Hz band power. Filters BGM/SFX
Pre-FFT energy check	Skips FFT for low-energy chunks (CPU savings). Threshold tuned to 3.0 (v6.2+)
Silence trimming	Cuts leading/trailing silence to reduce API payload
Audio normalization	DC offset removal + peak gain adjustment

Hallucination Detection

Feature	Description
Short text skip	Skips 1-word hallucinations <=3 chars
Repetition detection	Trims when same 3-char phrase repeats 8+ times
Prompt leak detection	Detects prompt leaks, skips

Note: Language-specific filters (Hangul, Cyrillic, Greek, Thai, Chinese, European languages, Nordic, Special Latin) are intentionally removed in the English version. English-version users may want to transcribe/translate those languages, so filtering them out would drop valid transcriptions. Only generic hallucination detection (short text, repetition, prompt leak) remains.

Filters & Quality Control

Feature	Description
NG word filter	200+ NG words pre-compiled as regex. Word boundary \b and substring match variants
Duplicate skip	Skips if >=85% similar to previous transcription
Translation loop detection	Trims when same phrase repeats in translation output
Translation safety refusal detection	Detects “I cannot provide” etc., uses original text
Translation failure detection	Detects Japanese returned as-is in ja_to_en, uses original text

Note on NG words: The NG word list includes 200+ words, many of which are Japanese-specific. If you are an English speaker using Guardrails ON (recommended), Cloudflare’s native content moderation will handle filtering – the built-in NG word list will have no visible effect on your transcription. If you use Guardrails OFF, the NG word list will still mask matching words, but most Japanese-specific entries will not match English speech. You can safely ignore the NG word list or customize NG_WORDS and NG_WORDS_BOUNDED in the script for your own language.

Recording & Pipeline

Feature	Description
InputStream continuous recording	PortAudio callback for gapless recording
Ring buffer	deque(maxlen=...) auto-discards old data on overrun
Non-blocking pipeline	Separate recording thread and API worker thread
HTTP session reuse	TCP/TLS handshake only once. Saves 100-300ms per cycle
Atomic write	Temp file then os.replace() prevents OBS read corruption
API retry	Both Whisper and translation API retry up to 2 times
Guardrails handling	Detects AI Gateway Guardrails blocks (code 2016/2017), skips

Modes & Settings

Feature	Description
Microphone mode	Picks up your voice. Dynamic noise gate enabled
Stream mode	Picks up game/stream audio. Higher VAD threshold
Bidirectional auto-detect	Auto-detects Japanese/English, decides translation direction. [JA]/[EN] tags
English-only mode	Picks up English only, passes through Japanese
Non-Japanese all mode	Picks up all non-Japanese audio, translates to Japanese
Latency modes	Choose 3s (low) / 5s (normal) / 8s (accuracy)
Genre hints	Game/chat/tech/music/kappa vocabulary hints for Whisper
Whisper model selection	Choose standard or high-accuracy model at startup (v6.2+)
15s auto-clear	Auto-clears subtitle after 15 seconds of silence
Environment variable	API token via CF_API_TOKEN environment variable

---

Requirements

• Python 3.10+
• pip packages: sounddevice, requests, numpy, scipy
• Cloudflare account (Free tier OK)
• OBS Studio (Text Source (GDI+) required)

pip install

pip install sounddevice requests numpy scipy

scipy is optional but recommended for higher-quality resampling.

---

Cloudflare Account Setup

Step 1: Create a Cloudflare Account

1. Sign up at Cloudflare
2. Log in to the Cloudflare Dashboard

Step 2: Get Your Account ID

1. Select your account in the dashboard sidebar
2. Copy the Account ID (32-character alphanumeric)
3. Set it in the script as ACCOUNT_ID

Step 3: Create an API Token

1. Go to API Tokens
2. Click Create Token
3. Select Create Custom Token
4. Set the following permissions:
   • Account > Workers AI – Read
   • Account > AI Gateway – Read, Edit
5. Continue to summary then Create Token
6. Copy the token (shown only once)

Step 4: Create an AI Gateway

1. Go to AI Gateway
2. Click Create Gateway
3. Enter a name (e.g., default)
4. Click Create

Auto-creation: Specifying default as the Gateway ID in API requests will auto-create the gateway on first request.

Guardrails: ON or OFF?

Audience	Guardrails	NG word list	Reason
English speakers	ON (recommended)	Not needed	Cloudflare’s native content moderation is high-accuracy for English. Simplest setup.
Japanese speakers	OFF (recommended)	Custom 200+ words	Neuron savings. Japanese NG expressions covered by custom filter.

Step 5: Enable AI Gateway Cache (Required – no Neuron savings without it)

1. Go to AI Gateway and select your gateway
2. Open Settings
3. Turn Cache Responses ON
4. Set default cache TTL to 300 seconds

Without this ON, caching won’t work. It won’t error, but Neuron savings will be zero. Make sure to turn it ON.

AI Gateway is completely free. Cache and logging features are free with no additional charges.

The script also specifies cache TTL via the cf-aig-cache-ttl header (TRANSLATE_CACHE_TTL = 300).

Step 6: Set Credentials in the Script

Using environment variables (recommended):

# Windows (Command Prompt)
set CF_API_TOKEN=your_token

# Windows (PowerShell)
$env:CF_API_TOKEN = "your_token"

# Mac / Linux
export CF_API_TOKEN=your_token

Or write directly into the script:

CLOUDFLARE_API_TOKEN = os.environ.get("CF_API_TOKEN", "cfat_paste_your_token_here")

Also set ACCOUNT_ID and GATEWAY_ID:

ACCOUNT_ID = "your_account_id"
GATEWAY_ID = "default"  # Gateway name you created

Step 7: Verify Workers AI Access

1. Go to Workers AI
2. Confirm the model list is displayed
3. Check Usage tab for Neuron consumption

Pricing (Don’t Worry)

Item	Cost
AI Gateway creation & usage	Free
AI Gateway cache feature	Free
AI Gateway logging	Free
Additional subscription	Not required

Workers AI (Whisper + LLM translation) pricing:

Plan	Free tier	Over limit	Auto-billing
Free (current)	10,000 Neurons/day free	Errors and stops	No
Paid ($5/mo)	10,000 Neurons/day free	$0.011 / 1,000 Neurons	Yes

You can use the Free plan. Up to 10,000 Neurons per day is free. No auto-billing on overage. It just errors and stops. Billing requires intentionally upgrading to Workers Paid ($5/mo).

---

Running the Script

python obs_subtitle_translator_0805_v6_3_en.py

At startup, you’ll be prompted to select:

1. Mode: Microphone (1) or Stream (2)
2. Language: Translation direction (1-5)
3. Device: Default (d) or list (l) or device number
4. Genre: Game/chat/tech/music/none/kappa (1-6)
5. Latency: Low (3s) / Normal (5s) / Accuracy (8s)
6. Whisper model: Standard (1) or High-accuracy (2)

Device Number Warning: The default device numbers (DEFAULT_MIC_DEVICE = 7, DEFAULT_GAME_DEVICE = 61) are specific to the author’s machine. Your device numbers will be different. Always enter l at the device selection prompt to list available devices and pick the correct one for your system.

---

OBS Configuration

1. Open OBS Studio
2. Sources then + then Text (GDI+)
3. Open the text source properties
4. Check Read from file
5. Browse and select output_mic.txt or output_game.txt
6. Set Encoding to UTF-8
7. Adjust font, size, and color

Subtitles update automatically. After 15 seconds of silence, the subtitle is cleared.

Mic Mode Usage Notes

• Stay silent for the first ~30 seconds after startup (noise floor learning in progress)
• Do not touch the mic during learning (noise floor will spike)
• Once [Noise Gate Confirmed] log appears, learning is done – OK to speak
• Learning with the same environment as streaming (with BGM) is recommended
• Press Ctrl+C to exit (output file will be cleared)

---

Use Cases

Example 1: Japanese Streamer to English Subtitles

Mode: 1 (Microphone)  Language: 1 (Japanese to English)  Latency: 2 (Normal 5s)

Example 2: English Game Stream to Japanese Subtitles

Mode: 2 (Stream)  Language: 1 (English to Japanese)  Latency: 3 (Accuracy 8s)

Example 3: Bidirectional Auto-Detect (Mixed JP/EN Stream)

Mode: 1 (Microphone)  Language: 3 (Auto-detect)  Latency: 2 (Normal 5s)

Example 4: English-Only Mode (Japanese Pass-Through)

Mode: 2 (Stream)  Language: 4 (English only, Japanese pass-through)

Example 5: Kappa Streaming (Vocabulary Hint)

“Kappa streaming” is a Japanese VTuber/streamer style where the streamer uses sentences ending with “-ppa” (a verbal tic). Genre 6 adds these patterns as Whisper vocabulary hints for better recognition. Non-Japanese speakers can ignore this option.

Mode: 1 (Microphone)  Language: 1 (Japanese to English)  Genre: 6 (Kappa)

Example 6: Tech Streaming (Technical Terms)

Mode: 1 (Microphone)  Language: 1 (Japanese to English)  Genre: 3 (Tech/Programming)

Example 7: Low Latency Mode

Mode: 1 (Microphone)  Language: 1 (Japanese to English)  Latency: 1 (Low 3s)

Example 8: Accuracy Priority for Noisy Environments

Mode: 2 (Stream)  Language: 1 (English to Japanese)  Latency: 3 (Accuracy 8s)

Example 9: High-Accuracy Whisper Model

Mode: 1 (Microphone)  Language: 1 (Japanese to English)  Whisper model: 2 (High-accuracy)

Example 10: French Stream to English Subtitles

Mode: 1 (Microphone)  Language: 2 (Speak English to Japanese)  Genre: 5 (None)

For French to English, use Whisper auto-detect mode (whisper_language = "auto") by modifying the script. The translation prompt can be changed to “Translate the following French to English” for best results.

Example 11: Korean Stream to English Subtitles

Mode: 1 (Microphone)  Language: 2 (Speak English to Japanese)  Genre: 5 (None)

Same as above – use Whisper auto-detect and modify the translation prompt. Language-specific hallucination filters are removed in the English version, so Korean speech will be transcribed correctly.

Example 12: Spanish Game Stream to English Subtitles

Mode: 2 (Stream)  Language: 1 (English to Japanese)  Genre: 5 (None)  Latency: 3 (Accuracy 8s)

For Spanish to English, use Whisper auto-detect and modify the translation direction. European language hallucination filters are removed in the English version, so Spanish speech will be transcribed correctly.

---

Whisper Model Selection

Model	Neurons/min	Accuracy	Free tier usage (approx.)
Standard (@cf/openai/whisper)	41.14	Medium	~8 hours
High-accuracy (whisper-large-v3-turbo)	46.63	High	~7 hours

Free tier is 10,000 Neurons/day. Streaming (with conversation gaps) gives 6-11 hours of usage. The high-accuracy model does not support temperature/prompt parameters and sends audio as base64 JSON.

---

Troubleshooting

Symptom	Cause	Fix
Not picking up voice	VAD precheck threshold too high	Lower VAD_PRECHECK_THRESHOLD from 3.0 to 2.0
Picking up too much noise	VAD precheck threshold too low	Raise VAD_PRECHECK_THRESHOLD from 3.0 to 5.0
Not picking up game audio	VAD threshold too high	Lower VAD_THRESHOLD_GAME from 400 to 300
Picking up too much game audio	VAD threshold too low	Raise VAD_THRESHOLD_GAME from 400 to 500
API error 401	Token permissions insufficient	Verify Workers AI Read + AI Gateway Read/Edit permissions
API timeout	Network latency	Raise API_TIMEOUT from 30 to 60
Subtitle garbled	OBS encoding	Set encoding to UTF-8 in Text Source (GDI+)
Nothing picked up at startup	Noise learning in progress (normal)	Wait 30 seconds. OK to speak after [Noise Gate Confirmed]
No [Cache Hit] shown	AI Gateway Cache Responses is OFF	Turn ON Cache Responses in dashboard, TTL 300s

---

Feature List

• Whisper speech recognition (standard / high-accuracy / English-only auto-select)
• LLM translation (llama-3.1-8b-instruct-fast)
• Dynamic noise gate (mic mode)
• Frequency VAD (with pre-FFT check)
• Hallucination detection (Repetition/Short text/Prompt leak – language-specific filters removed for multilingual support)
• NG word filter (word boundary support)
• Duplicate skip / Translation loop detection / Translation safety refusal detection
• 15-second auto-clear
• AI Gateway cache (TTL 300 seconds)
• Prompt Caching (x-session-affinity)
• Token usage display (Neuron monitoring)
• Whisper API / Translation API retry (up to 2 times)
• Auto-detect mode (bidirectional translation)

---

License

AI Model Licenses (as of 2026/08/05)

Model	Provider	License	Commercial Use
@cf/openai/whisper	OpenAI	MIT License	Allowed
@cf/openai/whisper-tiny-en	OpenAI	MIT License	Allowed
@cf/openai/whisper-large-v3-turbo	OpenAI	MIT License	Allowed
@cf/meta/llama-3.1-8b-instruct-fast	Meta	Llama Community License	Allowed (< 700M MAU)

Note: If you switch to different models (partner models, etc.), please check their respective terms of use and licenses.

Script License

Released under the MIT License. Feel free to use, modify, and distribute.

---

Disclaimer

• This script is provided “AS IS”, without any warranty
• Runs on the Cloudflare Workers AI free tier (10,000 Neurons/day), but may hit the limit depending on usage
• No auto-billing on overage (it just errors and stops)
• Billing requires intentionally upgrading to Workers Paid ($5/mo)
• AI speech recognition and translation are not perfect; misrecognition may occur
• The author is not responsible for any damages arising from use during streaming
• Cloudflare service specifications, pricing, and terms of use may change. Check the Cloudflare documentation for the latest information
• AI model license conditions may change. If you modify this script to use different models, be sure to check the applicable model licenses

---

Version History

v6.3 (0805)

• Version bump to align with Japanese version (0805)

v6.2 (0805)

• Translation API Prompt Caching (x-session-affinity) added
• VAD_PRECHECK_THRESHOLD adjusted 50.0 to 3.0 (hallucination reduction)
• VAD_THRESHOLD_MIC adjusted 50 to 100 (faint noise rejection)
• Kappa streaming prompt enhanced (added “-ppa” conversion phrases)
• Whisper model selection added (standard / high-accuracy)
• whisper-large-v3-turbo support (base64 JSON format, temperature/prompt not supported)
• Translation API token usage display added

v6.1 (0802)

• Pre-FFT energy check (CPU load reduction)
• Auto-select English Whisper model
• Translation API AI Gateway cache
• Pre-computed frequency mask
• Shortened VAD FFT window

v6.0

• InputStream + ring buffer (continuous recording)
• Non-blocking pipeline
• Pre-compiled NG word regex
• Translation API retry / Atomic write / Environment variable support
• HTTP session reuse / Latency mode selection
• Guardrails error handling
• Dynamic noise gate learning frame count

---

Differences from the Japanese Version

The English version (obs_subtitle_translator_0805_v6_3_en.py) is functionally identical to the Japanese version (obs_subtitle_translator_0805_v6_3.py) except for the following changes:

Removed Features

Feature	Japanese Version	English Version	Reason
Hangul detection	Detects Korean characters, skips	Removed	English-version users may want to transcribe Korean streams
Cyrillic detection	Detects Russian etc., skips	Removed	English-version users may want to transcribe Russian streams
Greek detection	Detects Greek, skips	Removed	English-version users may want to transcribe Greek streams
Thai detection	Detects Thai, skips	Removed	English-version users may want to transcribe Thai streams
Chinese detection	Detects simplified Chinese, skips	Removed	English-version users may want to transcribe Chinese streams
Special Latin detection	Detects Turkish characters, skips	Removed	English-version users may want to transcribe Turkish streams
Nordic detection	Detects Nordic characters, skips	Removed	English-version users may want to transcribe Nordic language streams
European language detection	Detects German/French/Swedish/Dutch/Spanish, skips	Removed	English-version users may want to transcribe European language streams
Kanji-only detection	Detects kanji-only text (<5 chars), skips	Removed	English-version users may want to transcribe kanji-only text

Kept Features

Feature	Status	Reason
Short text skip	Kept	Generic hallucination detection – not language-specific
Repetition detection	Kept	Generic hallucination detection – not language-specific
Prompt leak detection	Kept	Detects prompt leak – not language-specific
English hallucination skip (ja mode)	Kept	Only applies when whisper_language == "ja" – does not affect other modes
Translation failure detection (ja_to_en)	Kept	Only applies when direction == "ja_to_en" – does not affect other directions
NG word filter	Kept	All 200+ NG words kept, including Japanese NG words

Why These Changes Were Made

The Japanese version was built for a specific use case: a Japanese-speaking streamer who primarily speaks Japanese and occasionally needs English translation. In this context, Whisper sometimes produces hallucinations – random text in Korean, Chinese, German, French, etc. – when it misinterprets silence or noise. The language-specific filters were added to catch and discard these hallucinations, since the streamer would never actually speak those languages.

However, the English version targets international users who may speak any language:

• A French streamer wanting English subtitles would have their French speech incorrectly discarded by the European language filter
• A Korean streamer wanting English subtitles would have their Korean speech incorrectly discarded by the Hangul filter
• A Russian streamer wanting English subtitles would have their Russian speech incorrectly discarded by the Cyrillic filter
• A Chinese streamer wanting English subtitles would have their Chinese speech incorrectly discarded by the Chinese filter

Removing these filters ensures that valid speech in any language is never incorrectly discarded. The trade-off is that genuine Whisper hallucinations in those languages will no longer be filtered – but this is a better experience for multilingual users than having their actual speech dropped.

The NG word filter is kept because:

1. It only masks words (replaces with ***), it doesn’t discard the entire transcription
2. English speakers using Guardrails ON (recommended) will have Cloudflare’s native moderation handle this – the NG word list has no visible effect
3. Japanese speakers using Guardrails OFF still benefit from the 200+ Japanese NG words
4. Users can customize NG_WORDS and NG_WORDS_BOUNDED in the script for their own language

Summary

The Japanese version is optimized for Japanese streamers who primarily speak Japanese and need to filter out hallucinations in other languages. The English version is designed for international users who may transcribe and translate any language, so language-specific hallucination filters are removed to avoid dropping valid transcriptions.

---

References

• Cloudflare Workers AI
• Cloudflare AI Gateway
• Workers AI Model Catalog
• AI Gateway Caching
• Cloudflare API Token Creation

Use It :D