murmurai-core 1.0.2rc1

Modern speech recognition with word-level timestamps and speaker diarization. Fork of WhisperX with torch 2.6+, pyannote 4.x compatibility.

MurmurAI

Modern speech recognition with word-level timestamps and speaker diarization

---

MurmurAI is a fork of WhisperX with modern dependency support:

• PyTorch 2.6+ compatibility (weights_only patches)
• Pyannote 4.x support (token parameter migration)
• Torchaudio 2.9+ compatibility (audio backend fixes)
• Python 3.10-3.13 tested

Features

• Word-level timestamps via phoneme alignment
• Speaker diarization with pyannote.audio
• Batch inference for 70x realtime transcription
• VAD preprocessing (pyannote or silero)
• Multiple output formats: SRT, VTT, TXT, TSV, JSON

Installation

pip install murmurai-core

Or with uv:

uv add murmurai-core

Quick Start

Python API

import murmurai

# Load model
model = murmurai.load_model("large-v3-turbo", device="cuda", compute_type="float16")

# Transcribe
audio = murmurai.load_audio("audio.mp3")
result = model.transcribe(audio, batch_size=16)

# Align (word-level timestamps)
model_a, metadata = murmurai.load_align_model(language_code=result["language"], device="cuda")
result = murmurai.align(result["segments"], model_a, metadata, audio, device="cuda")

# Diarization (speaker labels)
from pyannote.audio import Pipeline
diarize_model = Pipeline.from_pretrained("pyannote/speaker-diarization-3.1", use_auth_token="YOUR_HF_TOKEN")
diarize_segments = diarize_model(audio)
result = murmurai.assign_word_speakers(diarize_segments, result)

CLI

Basic transcription:

murmurai-core audio.mp3

With specific model and language:

murmurai-core audio.mp3 --model large-v3-turbo --language en

With speaker diarization:

murmurai-core audio.mp3 --model large-v3-turbo --diarize --hf_token YOUR_HF_TOKEN

Output to specific format and directory:

murmurai-core audio.mp3 --output_format srt --output_dir ./transcripts

Run without installing (via uvx):

uvx murmurai-core audio.mp3 --model large-v3-turbo

Common options:

Option	Description	Default
--model	Whisper model (tiny, base, small, medium, large-v3, large-v3-turbo)	small
--language	Audio language (e.g., en, pt, es, fr)	auto-detect
--device	cuda or cpu	cuda
--output_format	srt, vtt, txt, tsv, json, all	all
--output_dir	Output directory	.
--diarize	Enable speaker diarization	off
--hf_token	HuggingFace token (required for diarization)	-
--batch_size	Batch size for inference	8
--compute_type	float16, float32, int8	float16

Requirements

• NVIDIA GPU with CUDA support (or CPU mode)
• HuggingFace token for diarization models

Accept the license at pyannote/speaker-diarization-3.1 before using diarization.

Migration from WhisperX

# Before
import whisperx

# After
import murmurai  # drop-in replacement

All APIs are identical. Just change the import.

Credits

MurmurAI builds on the excellent work of:

• WhisperX by Max Bain
• OpenAI Whisper
• Faster Whisper
• Pyannote

---

Made with ❤️ by Namastex Labs