speechlib 1.1.17

speechlib is a library that unifies speaker diarization, transcription and speaker recognition in a single pipeline to create transcripts for audio conversations with actual speaker names and time tags. This library also contain audio preprocessor functions.

Recall.ai - Meeting Transcription API

If you’re looking for a transcription API for meetings, consider checking out Recall.ai, an API that works with Zoom, Google Meet, Microsoft Teams, and more. Recall.ai diarizes by pulling the speaker data and separate audio streams from the meeting platforms, which means 100% accurate speaker diarization with actual speaker names.

Speechlib

Run your IDE as administrator

you will get following error if administrator permission is not there:

OSError: [WinError 1314] A required privilege is not held by the client

Requirements

• Python 3.8 or greater

GPU execution

GPU execution needs CUDA 11.

GPU execution requires the following NVIDIA libraries to be installed:

• cuBLAS for CUDA 11
• cuDNN 8 for CUDA 11

There are multiple ways to install these libraries. The recommended way is described in the official NVIDIA documentation, but we also suggest other installation methods below.

Google Colab:

on google colab run this to install CUDA dependencies:

!apt install libcublas11

You can see this example notebook

installation:

pip install speechlib
or
uv add speechlib

Dependencies:

dependencies = [
    "accelerate>=1.12.0",
    "assemblyai>=0.50.0",
    "faster-whisper>=1.2.1",
    "huggingface-hub==0.36.0",
    "numpy==1.26.4",
    "openai-whisper>=20250625",
    "pyannote-audio==3.4.0",
    "pydub>=0.25.1",
    "speechbrain==1.0.3",
    "torch==2.2.0",
    "torchaudio==2.2.0",
    "torchvision==0.17.0",
    "transformers>=4.57.6",
]

Introduction

This library does speaker diarization, speaker recognition, and transcription on a single wav file to provide a transcript with actual speaker names. This library will also return an array containing result information. ⚙

This library contains following audio preprocessing functions:

1. convert other audio formats to wav

2. convert stereo wav file to mono

3. re-encode the wav file to have 16-bit PCM encoding

Transcriptor method takes 7 arguments.

1. file to transcribe

2. log_folder to store transcription

3. language used for transcribing (language code is used)

4. model size ("tiny", "small", "medium", "large", "large-v1", "large-v2", "large-v3")

5. ACCESS_TOKEN: huggingface acccess token (also get permission to access pyannote/speaker-diarization@2.1)

6. voices_folder (contains speaker voice samples for speaker recognition)

7. quantization: this determine whether to use int8 quantization or not. Quantization may speed up the process but lower the accuracy.

voices_folder should contain subfolders named with speaker names. Each subfolder belongs to a speaker and it can contain many voice samples. This will be used for speaker recognition to identify the speaker.

if voices_folder is not provided then speaker tags will be arbitrary.

log_folder is to store the final transcript as a text file.

transcript will also indicate the timeframe in seconds where each speaker speaks.

Transcription example:

import os
from speechlib import Transcriptor

file = "obama_zach.wav"  # your audio file
voices_folder = "" # voices folder containing voice samples for recognition
language = "en"          # language code
log_folder = "logs"      # log folder for storing transcripts
modelSize = "tiny"     # size of model to be used [tiny, small, medium, large-v1, large-v2, large-v3]
quantization = False   # setting this 'True' may speed up the process but lower the accuracy
ACCESS_TOKEN = "huggingface api key" # get permission to access pyannote/speaker-diarization@2.1 on huggingface

# quantization only works on faster-whisper
transcriptor = Transcriptor(file, log_folder, language, modelSize, ACCESS_TOKEN, voices_folder, quantization)

# use normal whisper
res = transcriptor.whisper()

# use faster-whisper (simply faster)
res = transcriptor.faster_whisper()

# use a custom trained whisper model
res = transcriptor.custom_whisper("D:/whisper_tiny_model/tiny.pt")

# use a huggingface whisper model
res = transcriptor.huggingface_model("Jingmiao/whisper-small-chinese_base")

# use assembly ai model
res = transcriptor.assemby_ai_model("assemblyAI api key")

res --> [["start", "end", "text", "speaker"], ["start", "end", "text", "speaker"]...]

if you don't want speaker names: keep voices_folder as an empty string ""

start: starting time of speech in seconds
end: ending time of speech in seconds
text: transcribed text for speech during start and end
speaker: speaker of the text

voices folder structure:

voices_folder
|---> person1
|        |---> sample1.wav
|        |---> sample2.wav
|                ...
|
|---> person2
|        |---> sample1.wav
|        |---> sample2.wav
|                ...
|--> ...

supported language codes:

"af", "am", "ar", "as", "az", "ba", "be", "bg", "bn", "bo", "br", "bs", "ca", "cs", "cy", "da", "de", "el", "en", "es", "et", "eu", "fa", "fi", "fo", "fr", "gl", "gu", "ha", "haw", "he", "hi", "hr", "ht", "hu", "hy", "id", "is","it", "ja", "jw", "ka", "kk", "km", "kn", "ko", "la", "lb", "ln", "lo", "lt", "lv", "mg", "mi", "mk", "ml", "mn","mr", "ms", "mt", "my", "ne", "nl", "nn", "no", "oc", "pa", "pl", "ps", "pt", "ro", "ru", "sa", "sd", "si", "sk","sl", "sn", "so", "sq", "sr", "su", "sv", "sw", "ta", "te", "tg", "th", "tk", "tl", "tr", "tt", "uk", "ur", "uz","vi", "yi", "yo", "zh", "yue"

supported language names:

"Afrikaans", "Amharic", "Arabic", "Assamese", "Azerbaijani", "Bashkir", "Belarusian", "Bulgarian", "Bengali","Tibetan", "Breton", "Bosnian", "Catalan", "Czech", "Welsh", "Danish", "German", "Greek", "English", "Spanish","Estonian", "Basque", "Persian", "Finnish", "Faroese", "French", "Galician", "Gujarati", "Hausa", "Hawaiian","Hebrew", "Hindi", "Croatian", "Haitian", "Hungarian", "Armenian", "Indonesian", "Icelandic", "Italian", "Japanese","Javanese", "Georgian", "Kazakh", "Khmer", "Kannada", "Korean", "Latin", "Luxembourgish", "Lingala", "Lao","Lithuanian", "Latvian", "Malagasy", "Maori", "Macedonian", "Malayalam", "Mongolian", "Marathi", "Malay", "Maltese","Burmese", "Nepali", "Dutch", "Norwegian Nynorsk", "Norwegian", "Occitan", "Punjabi", "Polish", "Pashto","Portuguese", "Romanian", "Russian", "Sanskrit", "Sindhi", "Sinhalese", "Slovak", "Slovenian", "Shona", "Somali","Albanian", "Serbian", "Sundanese", "Swedish", "Swahili", "Tamil", "Telugu", "Tajik", "Thai", "Turkmen", "Tagalog","Turkish", "Tatar", "Ukrainian", "Urdu", "Uzbek", "Vietnamese", "Yiddish", "Yoruba", "Chinese", "Cantonese",

Audio preprocessing example:

from speechlib import PreProcessor

file = "obama1.mp3"
#initialize
prep = PreProcessor()
# convert mp3 to wav
wav_file = prep.convert_to_wav(file)

# convert wav file from stereo to mono
prep.convert_to_mono(wav_file)

# re-encode wav file to have 16-bit PCM encoding
prep.re_encode(wav_file)

Performance

These metrics are from Google Colab tests.
These metrics do not take into account model download times.
These metrics are done without quantization enabled.
(quantization will make this even faster)

metrics for faster-whisper "tiny" model:
    on gpu:
        audio name: obama_zach.wav
        duration: 6 min 36 s
        diarization time: 24s
        speaker recognition time: 10s
        transcription time: 64s


metrics for faster-whisper "small" model:
    on gpu:
        audio name: obama_zach.wav
        duration: 6 min 36 s
        diarization time: 24s
        speaker recognition time: 10s
        transcription time: 95s


metrics for faster-whisper "medium" model:
    on gpu:
        audio name: obama_zach.wav
        duration: 6 min 36 s
        diarization time: 24s
        speaker recognition time: 10s
        transcription time: 193s


metrics for faster-whisper "large" model:
    on gpu:
        audio name: obama_zach.wav
        duration: 6 min 36 s
        diarization time: 24s
        speaker recognition time: 10s
        transcription time: 343s

Citation

If you use Speechlib in your research or project, please cite it as:

@software{speechlib,
  author       = {NavodPeiris},
  title        = {Speechlib: Speaker Diarization, Recognition, and Transcription in a Single Pipeline},
  year         = {2024},
  publisher    = {GitHub},
  url          = {https://github.com/NavodPeiris/speechlib}
}

Sponsorship

If you find Speechlib useful, please consider supporting its development:

• GitHub Sponsors

Your support helps maintain and improve the library. Thank you!