feat: update transcription engine (#16)

* feat: update transcription engine

* fix engine values

Dockerfile

@@ -11,7 +11,7 @@ RUN apt-get install -y ffmpeg nodejs=18.* \
     && rm -rf /var/lib/apt/lists/*
 
 # install python dependencies
-RUN pip install -U openai-whisper
+RUN pip install -U openai-whisper whisper-ctranslate2
 
 # add source files
 COPY src /app

Makefile

@@ -32,7 +32,7 @@ test-clean:
 whisper:
 	docker run -it --rm \
 		-v ${DIR}:/app \
-		-v ${DIR}/.whisper:/root/.cache/whisper \
+		-v ${DIR}/.whisper:/app/.whisper \
 		${IMAGE_TAG} \
 		whisper ${VIDEO_FILE} \
 			--model ${MODEL} \
@@ -43,11 +43,12 @@ whisper:
 node:
 	docker run -it --rm \
 		-v ${DIR}:/app \
-		-v ${DIR}/.whisper:/root/.cache/whisper \
+		-v ${DIR}/.whisper:/app/.whisper \
 		${IMAGE_TAG} \
 		node /app/src/cut-video.js -t test -v test.mp4
 
 bash:
 	docker run -it --rm \
 		-v ${DIR}:/data \
+		-v ${DIR}/.whisper:/app/.whisper \
 		${IMAGE_TAG} /bin/bash

README.md

@@ -35,6 +35,7 @@ docker run --rm -it \
 - `--input` - path to video file
 - `--model` - whisper model name - `tiny`, `tiny.en` (default), `base`, `base.en`, `small`, `small.en`, `medium`, `medium.sm`, `large`. View [official docs](https://github.com/openai/whisper#available-models-and-languages) for break down of model size and performance
 - `--language` - language code. Typically improves in transcription to set language instead of allowing Whisper to auto-detect.
+- `--engine` - transcription engine `whisper-ctranslate2` (default) or `whisper`. `whisper` is likely to be removed in the near future if/when `whisper-ctranslate2` proves to be just as good but 4x faster
 
 ### Known Issues
 - `Error: Command "whisper" exited with code null` - this is likely caused by the container needing more allocated memory. Allocating at least 4 GB memory for the `small.en` usually resolved the issue but your mileage may vary.
@@ -47,7 +48,7 @@ Allows you to manually create a list of timestamps to cut the video.
 
 Usage:
 ```shell
-docker run --rm -it -v $(pwd):/data video-swear-jar \
+docker run --rm -it -v $(pwd):/data video-swear-jar:v1 \
   cut-video --timestamp timestamps.txt --video video.mkv
 ```
 
@@ -60,7 +61,7 @@ This is the `whisper` CLI if you need to further customize the command. Visit ht
 
 Usage:
 ```shell
-docker run --rm -it -v $(pwd):/data video-swear-jar \
+docker run --rm -it -v $(pwd):/data video-swear-jar:v1 \
   whisper my-video.mp4 \
     --model tiny.en \
     --language en \
@@ -71,7 +72,7 @@ docker run --rm -it -v $(pwd):/data video-swear-jar \
 ### ffmpeg
 Usage:
 ```shell
-docker run --rm -it -v $(pwd):/data video-swear-jar \
+docker run --rm -it -v $(pwd):/data video-swear-jar:v1 \
   ffmpeg -i input.mp4 output.avi
 ```
 

docs/notes.md

@@ -7,4 +7,9 @@ ffmpeg \
   -i input.mkv \
   -vcodec copy \
   -acodec copy output.mkv
-```
+```
+
+## Whisper Research
+- [Making OpenAI Whisper faster](https://nikolas.blog/making-openai-whisper-faster/)
+- https://github.com/guillaumekln/faster-whisper - Faster Whisper transcription with CTranslate2
+- https://github.com/Softcatala/whisper-ctranslate2 - Whisper command line client compatible with original OpenAI client based on CTranslate2

src/clean.js

@@ -4,18 +4,24 @@ const log = require('./log')
 const utils = require('./utils')
 const video = require('./video')
 
-const argv = yargs.usage('clean')
+const argv = yargs.usage('clean-fast')
   .options({
     input: {
       description: 'Input video filename',
       demandOption: true,
       alias: 'i'
     },
+    engine: {
+      description: 'Transcription engine',
+      alias: 'e',
+      default: 'whisper-ctranslate2',
+      choices: ['whisper', 'whisper-ctranslate2']
+    },
     model: {
       description: 'Whisper model name',
       alias: 'm',
       default: 'tiny.en',
-      choices: ['tiny.en', 'tiny', 'base.en', 'base', 'small.en', 'small']
+      choices: ['tiny.en', 'tiny', 'base.en', 'base', 'small.en', 'small', 'medium.en', 'medium', 'large-v1', 'large-v2', 'large']
     },
     language: {
       description: 'Video file language',
@@ -41,7 +47,7 @@ const run = async () => {
   try {
     log.info('[1 of 4] Starting video transcribe...')
     const { model, language } = argv
-    await video.transcribe({ inputFile: paths.inputFile, model, language, outputDir: argv['output-dir'] })
+    await video.transcribe({ engine: argv.engine, inputFile: paths.inputFile, model, language, outputDir: argv['output-dir'] })
   } catch (err) {
     log.error(`Unable to transcribe ${paths.inputFile}`, err)
     throw err

src/video.js

@@ -2,17 +2,27 @@ const fs = require('fs')
 const swearWords = require('./swear-words.json')
 const utils = require('./utils')
 
-const transcribe = async ({ inputFile, model = 'tiny.en', language = 'en', outputDir = '.' }) => {
+const transcribe = async ({ engine = 'whisper-ctranslate2', inputFile, model = 'tiny.en', language = 'en', outputDir = '.' }) => {
   const args = [
     inputFile,
     '--model', model,
     '--model_dir', '/app/.whisper',
     '--language', language,
     '--output_format', 'json',
-    '--output_dir', outputDir,
-    '--fp16', 'False' // TODO: make CLI argument to use GPU
+    '--output_dir', outputDir
   ]
-  await utils.asyncSpawn('whisper', args)
+
+  // engine specific args
+  switch (engine) {
+    case 'whisper':
+      args.push('--fp16', 'False')
+      break
+    case 'whisper-ctranslate2':
+      args.push('--compute_type', 'int8')
+      break
+  }
+
+  await utils.asyncSpawn(engine, args)
 }
 
 const cut = async ({ cutFile, outputFile }) => {