diarizer-lite

A Lightweight speaker diarization library using stereo audio energy for timestamped speech-to-text (STT) transcripts.

What this package does

Most modern STT systems such as - Whisper, Whisper-faster, Deepgram, AssemblyAI, AWS Transcribe, Google STT, etc... can output timestamped segments of text, but they either:

• don’t do speaker diarization at all, or
• require cloud billing for diarization, or
• produce overly granular segments (short “speaker turns”), or
• lack merging into coherent speaker turns for downstream use

diarizer-lite fills this gap by performing lightweight speaker assignment using stereo audio channel energy (left/right RMS) to determine who spoke when.

Unlike ML-based diarization (e.g. Pyannote, embeddings, VAD models), this method:

• does not require GPUs
• does not require machine learning
• runs offline
• works with raw stereo call recordings with any format (mp3, wav, etc.)
• merges consecutive segments into human-readable speaker turns

This makes it ideal for:

• call center analytics
• customer ↔ agent conversations
• post-processing Whisper transcripts
• LLM conversation input formatting (context engineering)
• summarization pipelines
• sentiment/emotion analysis
• compliance/regulatory auditing

Why this matters

Large Language Models (LLMs) can summarize, analyze sentiment, and extract insights from transcripts — but only if speaker turns are well structured.

Example:

“Okay. Okay. Okay.”
doesn’t help a summarizer as separate lines, but merged as one speaker turn it becomes useful context.

diarizer-lite converts fragmented timestamped segments into proper conversational turns, enabling LLMs to better understand who said what and when.

Supported Inputs

• Audio: .wav, .mp3, .flac (stereo recommended)
• Segments: list of {start, end, text} dicts from STT systems

Compatible with outputs from:

• Whisper / Faster-Whisper
• Deepgram
• AssemblyAI
• AWS Transcribe (post-word-grouping)
• Google Speech-to-Text (post-word-grouping)
• Riva / Vosk / Coqui
• Any STT that outputs timestamps

Installation

pip install diarizer-lite

Usage

from diarizer_lite import Diarizer

d = Diarizer()

diarized = d.diarize_segments(
    audio_file="call.wav",
    segments=segments
)

Before / After Example

Input (timestamped STT segments)

segments = [
    {"start": 0.0, "end": 4.0, "text": "Hi, I wanted to report an issue with my ride this morning."},
    {"start": 4.0, "end": 7.0, "text": "Sure, could you tell me what happened?"},
    {"start": 7.0, "end": 13.0, "text": "Yeah, driver was polite but car wasn't clean and smelled weird."},
    {"start": 13.0, "end": 17.0, "text": "I'm sorry to hear that. Anything else?"},
    {"start": 17.0, "end": 22.0, "text": "Also he took a longer route even after I gave the correct address."},
    {"start": 22.0, "end": 25.0, "text": "Got it. We will look into this and report back."}
]

Output (diarized speaker turns)

[
  {"start": 0.0, "end": 4.0, "speaker": "Speaker0",
   "text": "Hi, I wanted to report an issue with my ride this morning."},

  {"start": 4.0, "end": 7.0, "speaker": "Speaker1",
   "text": "Sure, could you tell me what happened?"},

  {"start": 7.0, "end": 13.0, "speaker": "Speaker0",
   "text": "Yeah, driver was polite but car wasn't clean and smelled weird."},

  {"start": 13.0, "end": 17.0, "speaker": "Speaker1",
   "text": "I'm sorry to hear that. Anything else?"},

  {"start": 17.0, "end": 25.0, "speaker": "Speaker0",
   "text": "Also he took a longer route even after I gave the correct address. Got it. We will look into this and report back."}
]

Mono Fallback

If audio is mono, diarizer-lite cannot infer speakers and returns:

speaker="Unknown"

License

MIT License