LLM ASR

package.json

@@ -1,6 +1,6 @@
 {
 	"name": "@ztimson/ai-utils",
-	"version": "0.8.0",
+	"version": "0.8.1",
 	"description": "AI Utility library",
 	"author": "Zak Timson",
 	"license": "MIT",

src/audio.ts

@@ -35,6 +35,121 @@ print(json.dumps(segments))
 `;
 	}
 
+	private async addPunctuation(timestampData: any, llm?: boolean, cadence = 150): Promise<string> {
+		const countSyllables = (word: string): number => {
+			word = word.toLowerCase().replace(/[^a-z]/g, '');
+			if(word.length <= 3) return 1;
+			const matches = word.match(/[aeiouy]+/g);
+			let count = matches ? matches.length : 1;
+			if(word.endsWith('e')) count--;
+			return Math.max(1, count);
+		};
+
+		let result = '';
+		timestampData.transcription.filter((word, i) => {
+			let skip = false;
+			const prevWord = timestampData.transcription[i - 1];
+			const nextWord = timestampData.transcription[i + 1];
+			if(!word.text && nextWord) {
+				nextWord.offsets.from = word.offsets.from;
+				nextWord.timestamps.from = word.offsets.from;
+			} else if(word.text && word.text[0] != ' ' && prevWord) {
+				prevWord.offsets.to = word.offsets.to;
+				prevWord.timestamps.to = word.timestamps.to;
+				prevWord.text += word.text;
+				skip = true;
+			}
+			return !!word.text && !skip;
+		}).forEach((word: any) => {
+			const capital = /^[A-Z]/.test(word.text.trim());
+			const length = word.offsets.to - word.offsets.from;
+			const syllables = countSyllables(word.text.trim());
+			const expected = syllables * cadence;
+			if(capital && length > expected * 2 && word.text[0] == ' ') result += '.';
+			result += word.text;
+		});
+		if(!llm) return result.trim();
+		return this.ai.language.ask(result, {
+			system: 'Remove any misplaced punctuation from the following ASR transcript using the replace tool. Avoid modifying words unless there is an obvious typo',
+			temperature: 0.1,
+			tools: [{
+				name: 'replace',
+				description: 'Use find and replace to fix errors',
+				args: {
+					find: {type: 'string', description: 'Text to find', required: true},
+					replace: {type: 'string', description: 'Text to replace', required: true}
+				},
+				fn: (args) => result = result.replace(args.find, args.replace)
+			}]
+		}).then(() => result);
+	}
+
+	private async diarizeTranscript(timestampData: any, speakers: any[], llm: boolean): Promise<string> {
+		const speakerMap = new Map();
+		let speakerCount = 0;
+		speakers.forEach((seg: any) => {
+			if(!speakerMap.has(seg.speaker)) speakerMap.set(seg.speaker, ++speakerCount);
+		});
+
+		const punctuatedText = await this.addPunctuation(timestampData, llm);
+		const sentences = punctuatedText.match(/[^.!?]+[.!?]+/g) || [punctuatedText];
+		const words = timestampData.transcription.filter((w: any) => w.text.trim());
+
+		// Assign speaker to each sentence
+		const sentencesWithSpeakers = sentences.map(sentence => {
+			sentence = sentence.trim();
+			if(!sentence) return null;
+
+			const sentenceWords = sentence.toLowerCase().replace(/[^\w\s]/g, '').split(/\s+/);
+			const speakerWordCount = new Map<number, number>();
+
+			sentenceWords.forEach(sw => {
+				const word = words.find((w: any) => sw === w.text.trim().toLowerCase().replace(/[^\w]/g, ''));
+				if(!word) return;
+
+				const wordTime = word.offsets.from / 1000;
+				const speaker = speakers.find((seg: any) => wordTime >= seg.start && wordTime <= seg.end);
+				if(speaker) {
+					const spkNum = speakerMap.get(speaker.speaker);
+					speakerWordCount.set(spkNum, (speakerWordCount.get(spkNum) || 0) + 1);
+				}
+			});
+
+			let bestSpeaker = 1;
+			let maxWords = 0;
+			speakerWordCount.forEach((count, speaker) => {
+				if(count > maxWords) {
+					maxWords = count;
+					bestSpeaker = speaker;
+				}
+			});
+
+			return {speaker: bestSpeaker, text: sentence};
+		}).filter(s => s !== null);
+
+		// Merge adjacent sentences from same speaker
+		const merged: Array<{speaker: number, text: string}> = [];
+		sentencesWithSpeakers.forEach(item => {
+			const last = merged[merged.length - 1];
+			if(last && last.speaker === item.speaker) {
+				last.text += ' ' + item.text;
+			} else {
+				merged.push({...item});
+			}
+		});
+
+		let transcript = merged.map(item => `[Speaker ${item.speaker}]: ${item.text}`).join('\n').trim();
+		if(!llm) return transcript;
+		let chunks = this.ai.language.chunk(transcript, 500, 0);
+		if(chunks.length > 4) chunks = [...chunks.slice(0, 3), <string>chunks.at(-1)];
+		const names = await this.ai.language.json(chunks.join('\n'), '{1: "Detected Name", 2: "Second Name"}', {
+			system: 'Use the following transcript to identify speakers. Only identify speakers you are positive about, dont mention speakers you are unsure about in your response',
+			temperature: 0.1,
+		});
+		Object.entries(names).forEach(([speaker, name]) => transcript = transcript.replaceAll(`[Speaker ${speaker}]`, `[${name}]`));
+		return transcript;
+	}
+
 	private runAsr(file: string, opts: {model?: string, diarization?: boolean} = {}): AbortablePromise<any> {
 		let proc: any;
 		const p = new Promise<any>((resolve, reject) => {
@@ -111,102 +226,28 @@ print(json.dumps(segments))
 		return <any>Object.assign(p, {abort});
 	}
 
-	private async combineSpeakerTranscript(punctuatedText: string, timestampData: any, speakers: any[]): Promise<string> {
-		const speakerMap = new Map();
-		let speakerCount = 0;
-		speakers.forEach((seg: any) => {
-			if(!speakerMap.has(seg.speaker)) speakerMap.set(seg.speaker, ++speakerCount);
-		});
-
-		const sentences = punctuatedText.match(/[^.!?]+[.!?]+/g) || [punctuatedText];
-		const lines: string[] = [];
-
-		sentences.forEach(sentence => {
-			sentence = sentence.trim();
-			if(!sentence) return;
-
-			const words = sentence.toLowerCase().replace(/[^\w\s]/g, '').split(/\s+/);
-			let startTime = Infinity, endTime = 0;
-			const wordTimings: {start: number, end: number}[] = [];
-
-			timestampData.transcription.forEach((word: any) => {
-				const wordText = word.text.trim().toLowerCase();
-				if(words.some(w => wordText.includes(w))) {
-					const start = word.offsets.from / 1000;
-					const end = word.offsets.to / 1000;
-					wordTimings.push({start, end});
-					if(start < startTime) startTime = start;
-					if(end > endTime) endTime = end;
-				}
-			});
-
-			if(startTime === Infinity) return;
-
-			// Weight by word-level overlap instead of sentence span
-			const speakerScores = new Map<number, number>();
-
-			wordTimings.forEach(wt => {
-				speakers.forEach((seg: any) => {
-					const overlap = Math.max(0, Math.min(wt.end, seg.end) - Math.max(wt.start, seg.start));
-					const duration = wt.end - wt.start;
-					if(duration > 0) {
-						const score = overlap / duration; // % of word covered
-						const spkNum = speakerMap.get(seg.speaker);
-						speakerScores.set(spkNum, (speakerScores.get(spkNum) || 0) + score);
-					}
-				});
-			});
-
-			let bestSpeaker = 1;
-			let maxScore = 0;
-			speakerScores.forEach((score, speaker) => {
-				if(score > maxScore) {
-					maxScore = score;
-					bestSpeaker = speaker;
-				}
-			});
-
-			lines.push(`[Speaker ${bestSpeaker}]: ${sentence}`);
-		});
-
-		return lines.join('\n').trim();
-	}
-
-	asr(file: string, options: { model?: string; diarization?: boolean | 'id' } = {}): AbortablePromise<string | null> {
+	asr(file: string, options: { model?: string; diarization?: boolean | 'llm' } = {}): AbortablePromise<string | null> {
 		if(!this.ai.options.whisper) throw new Error('Whisper not configured');
 
 		const tmp = join(mkdtempSync(join(tmpdir(), 'audio-')), 'converted.wav');
 		execSync(`ffmpeg -i "${file}" -ar 16000 -ac 1 -f wav "${tmp}"`, { stdio: 'ignore' });
 		const clean = () => fs.rm(Path.dirname(tmp), {recursive: true, force: true}).catch(() => {});
-		const transcript = this.runAsr(tmp, {model: options.model, diarization: false});
-		const timestamps: any = !options.diarization ? Promise.resolve(null) : this.runAsr(tmp, {model: options.model, diarization: true});
-		const diarization: any = !options.diarization ? Promise.resolve(null) : this.runDiarization(tmp);
+
+		if(!options.diarization) return this.runAsr(tmp, {model: options.model});
+		const timestamps = this.runAsr(tmp, {model: options.model, diarization: true});
+		const diarization = this.runDiarization(tmp);
 		let aborted = false, abort = () => {
 			aborted = true;
-			transcript.abort();
-			timestamps?.abort?.();
-			diarization?.abort?.();
+			timestamps.abort();
+			diarization.abort();
 			clean();
 		};
 
-		const response = Promise.allSettled([transcript, timestamps, diarization]).then(async ([t, ts, d]) => {
-			if(t.status == 'rejected') throw new Error('Whisper.cpp punctuated:\n' + t.reason);
+		const response = Promise.allSettled([timestamps, diarization]).then(async ([ts, d]) => {
 			if(ts.status == 'rejected') throw new Error('Whisper.cpp timestamps:\n' + ts.reason);
 			if(d.status == 'rejected') throw new Error('Pyannote:\n' + d.reason);
-			if(aborted || !options.diarization) return t.value;
-
-			let transcript = await this.combineSpeakerTranscript(t.value, ts.value, d.value);
-			if(!aborted && options.diarization === 'id') {
-				if(!this.ai.language.defaultModel) throw new Error('Configure an LLM for advanced ASR speaker detection');
-				let chunks = this.ai.language.chunk(transcript, 500, 0);
-				if(chunks.length > 4) chunks = [...chunks.slice(0, 3), <string>chunks.at(-1)];
-				const names = await this.ai.language.json(chunks.join('\n'), '{1: "Detected Name", 2: "Second Name"}', {
-					system: 'Use the following transcript to identify speakers. Only identify speakers you are positive about, dont mention speakers you are unsure about in your response',
-					temperature: 0.1,
-				});
-				Object.entries(names).forEach(([speaker, name]) => transcript = transcript.replaceAll(`[Speaker ${speaker}]`, `[${name}]`));
-			}
-			return transcript;
+			if(aborted || !options.diarization) return ts.value;
+			return this.diarizeTranscript(ts.value, d.value, options.diarization == 'llm');
 		}).finally(() => clean());
 		return <any>Object.assign(response, {abort});
 	}

src/llm.ts

@@ -145,7 +145,7 @@ class LLM {
 
 			// Handle compression and memory extraction
 			if(options.compress || options.memory) {
-				let compressed = null;
+				let compressed: any = null;
 				if(options.compress) {
 					compressed = await this.ai.language.compressHistory(options.history, options.compress.max, options.compress.min, options);
 					options.history.splice(0, options.history.length, ...compressed.history);
@@ -164,6 +164,15 @@ class LLM {
 		}), {abort});
 	}
 
+	async code(message: string, options?: LLMRequest): Promise<any> {
+		const resp = await this.ask(message, {...options, system: [
+			options?.system,
+			'Return your response in a code block'
+		].filter(t => !!t).join(('\n'))});
+		const codeBlock = /```(?:.+)?\s*([\s\S]*?)```/.exec(resp);
+		return codeBlock ? codeBlock[1].trim() : null;
+	}
+
 	/**
 	 * Compress chat history to reduce context size
 	 * @param {LLMMessage[]} history Chatlog that will be compressed
@@ -343,14 +352,11 @@ class LLM {
 	 * @returns {Promise<{} | {} | RegExpExecArray | null>}
 	 */
 	async json(text: string, schema: string, options?: LLMRequest): Promise<any> {
-		let resp = await this.ask(text, {...options, system: (options?.system ? `${options.system}\n` : '') + `Only respond using a JSON code block matching this schema:
-\`\`\`json
-${schema}
-\`\`\``});
-		if(!resp) return {};
-		const codeBlock = /```(?:.+)?\s*([\s\S]*?)```/.exec(resp);
-		const jsonStr = codeBlock ? codeBlock[1].trim() : resp;
-		return JSONAttemptParse(jsonStr, {});
+		const code = await this.code(text, {...options, system: [
+			options?.system,
+			`Only respond using JSON matching this schema:\n\`\`\`json\n${schema}\n\`\`\``
+		].filter(t => !!t).join('\n')});
+		return code ? JSONAttemptParse(code, {}) : null;
 	}
 
 	/**

tsconfig.json

@@ -15,6 +15,7 @@
 		"noEmit": true,
 
 		/* Linting */
-		"strict": true
+		"strict": true,
+		"noImplicitAny": false
 	}
 }