Video Subtitle Generation using Whisper and OpenVINO™

This Jupyter notebook can be launched on-line, opening an interactive environment in a browser window. You can also make a local installation. Choose one of the following options:

Google ColabGithub

Whisper is an automatic speech recognition (ASR) system trained on 680,000 hours of multilingual and multitask supervised data collected from the web. It is a multi-task model that can perform multilingual speech recognition as well as speech translation and language identification.



You can find more information about this model in the research paper, OpenAI blog, model card and GitHub repository.

In this notebook, we will use Whisper with OpenVINO to generate subtitles in a sample video. Notebook contains the following steps:

  1. Download the model.

  2. Instantiate the PyTorch model pipeline.

  3. Convert model to OpenVINO IR, using model conversion API.

  4. Run the Whisper pipeline with OpenVINO models.

Table of contents:


Install dependencies.

%pip install -q "openvino>=2023.1.0"
%pip install -q "python-ffmpeg<=1.0.16" moviepy transformers --extra-index-url
%pip install -q "git+"
%pip install -q  gradio
%pip install -q "openai-whisper==20231117" --extra-index-url

Instantiate model

Whisper is a Transformer based encoder-decoder model, also referred to as a sequence-to-sequence model. It maps a sequence of audio spectrogram features to a sequence of text tokens. First, the raw audio inputs are converted to a log-Mel spectrogram by action of the feature extractor. Then, the Transformer encoder encodes the spectrogram to form a sequence of encoder hidden states. Finally, the decoder autoregressively predicts text tokens, conditional on both the previous tokens and the encoder hidden states.

You can see the model architecture in the diagram below:



There are several models of different sizes and capabilities trained by the authors of the model. In this tutorial, we will use the base model, but the same actions are also applicable to other models from Whisper family.

from whisper import _MODELS
import ipywidgets as widgets

model_id = widgets.Dropdown(

Dropdown(description='Model:', index=9, options=('tiny.en', 'tiny', 'base.en', 'base', 'small.en', 'small', 'm…
import whisper

model = whisper.load_model(model_id.value, "cpu")

Convert model to OpenVINO Intermediate Representation (IR) format.

For best results with OpenVINO, it is recommended to convert the model to OpenVINO IR format. We need to provide initialized model object and example of inputs for shape inference. We will use ov.convert_model functionality to convert models. The ov.convert_model Python function returns an OpenVINO model ready to load on device and start making predictions. We can save it on disk for next usage with ov.save_model.

Convert Whisper Encoder to OpenVINO IR

from pathlib import Path

WHISPER_ENCODER_OV = Path(f"whisper_{model_id.value}_encoder.xml")
WHISPER_DECODER_OV = Path(f"whisper_{model_id.value}_decoder.xml")
import torch
import openvino as ov

mel = torch.zeros((1, 80 if 'v3' not in model_id.value else 128, 3000))
audio_features = model.encoder(mel)
if not WHISPER_ENCODER_OV.exists():
    encoder_model = ov.convert_model(model.encoder, example_input=mel)
    ov.save_model(encoder_model, WHISPER_ENCODER_OV)

Convert Whisper decoder to OpenVINO IR

To reduce computational complexity, the decoder uses cached key/value projections in attention modules from the previous steps. We need to modify this process for correct tracing.

import torch
from typing import Optional, Tuple
from functools import partial

def attention_forward(
        x: torch.Tensor,
        xa: Optional[torch.Tensor] = None,
        mask: Optional[torch.Tensor] = None,
        kv_cache: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,
    Override for forward method of decoder attention module with storing cache values explicitly.
      attention_module: current attention module
      x: input token ids.
      xa: input audio features (Optional).
      mask: mask for applying attention (Optional).
      kv_cache: dictionary with cached key values for attention modules.
      idx: idx for search in kv_cache.
      attention module output tensor
      updated kv_cache
    q = attention_module.query(x)

    if xa is None:
        # hooks, if installed (i.e. kv_cache is not None), will prepend the cached kv tensors;
        # otherwise, perform key/value projections for self- or cross-attention as usual.
        k = attention_module.key(x)
        v = attention_module.value(x)
        if kv_cache is not None:
            k =[0], k), dim=1)
            v =[1], v), dim=1)
        kv_cache_new = (k, v)
        # for cross-attention, calculate keys and values once and reuse in subsequent calls.
        k = attention_module.key(xa)
        v = attention_module.value(xa)
        kv_cache_new = (None, None)

    wv, qk = attention_module.qkv_attention(q, k, v, mask)
    return attention_module.out(wv), kv_cache_new

def block_forward(
    x: torch.Tensor,
    xa: Optional[torch.Tensor] = None,
    mask: Optional[torch.Tensor] = None,
    kv_cache: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,
    Override for residual block forward method for providing kv_cache to attention module.
        residual_block: current residual block.
        x: input token_ids.
        xa: input audio features (Optional).
        mask: attention mask (Optional).
        kv_cache: cache for storing attention key values.
        x: residual block output
        kv_cache: updated kv_cache

    x0, kv_cache = residual_block.attn(residual_block.attn_ln(
        x), mask=mask, kv_cache=kv_cache)
    x = x + x0
    if residual_block.cross_attn:
        x1, _ = residual_block.cross_attn(
            residual_block.cross_attn_ln(x), xa)
        x = x + x1
    x = x + residual_block.mlp(residual_block.mlp_ln(x))
    return x, kv_cache

# update forward functions
for idx, block in enumerate(model.decoder.blocks):
    block.forward = partial(block_forward, block)
    block.attn.forward = partial(attention_forward, block.attn)
    if block.cross_attn:
        block.cross_attn.forward = partial(attention_forward, block.cross_attn)

def decoder_forward(decoder, x: torch.Tensor, xa: torch.Tensor, kv_cache: Optional[Tuple[Tuple[torch.Tensor, torch.Tensor]]] = None):
    Override for decoder forward method.
      x: torch.LongTensor, shape = (batch_size, <= n_ctx) the text tokens
      xa: torch.Tensor, shape = (batch_size, n_mels, n_audio_ctx)
           the encoded audio features to be attended on
      kv_cache: Dict[str, torch.Tensor], attention modules hidden states cache from previous steps
    if kv_cache is not None:
        offset = kv_cache[0][0].shape[1]
        offset = 0
        kv_cache = [None for _ in range(len(decoder.blocks))]
    x = decoder.token_embedding(
        x) + decoder.positional_embedding[offset: offset + x.shape[-1]]
    x =
    kv_cache_upd = []

    for block, kv_block_cache in zip(decoder.blocks, kv_cache):
        x, kv_block_cache_upd = block(x, xa, mask=decoder.mask, kv_cache=kv_block_cache)

    x = decoder.ln(x)
    logits = (
        x @ torch.transpose(, 1, 0)).float()

    return logits, tuple(kv_cache_upd)

# override decoder forward
model.decoder.forward = partial(decoder_forward, model.decoder)
tokens = torch.ones((5, 3), dtype=torch.int64)
logits, kv_cache = model.decoder(tokens, audio_features, kv_cache=None)

tokens = torch.ones((5, 1), dtype=torch.int64)

if not WHISPER_DECODER_OV.exists():
    decoder_model = ov.convert_model(model.decoder, example_input=(tokens, audio_features, kv_cache))
    ov.save_model(decoder_model, WHISPER_DECODER_OV)

The decoder model autoregressively predicts the next token guided by encoder hidden states and previously predicted sequence. This means that the shape of inputs which depends on the previous step (inputs for tokens and attention hidden states from previous step) are dynamic. For efficient utilization of memory, you define an upper bound for dynamic input shapes.

Prepare inference pipeline

The image below illustrates the pipeline of video transcribing using the Whisper model.



To run the PyTorch Whisper model, we just need to call the model.transcribe(audio, **parameters) function. We will try to reuse original model pipeline for audio transcribing after replacing the original models with OpenVINO IR versions.

### Select inference device

select device from dropdown list for running inference using OpenVINO

core = ov.Core()
import ipywidgets as widgets

device = widgets.Dropdown(
    options=core.available_devices + ["AUTO"],

Dropdown(description='Device:', index=2, options=('CPU', 'GPU', 'AUTO'), value='AUTO')
from utils import patch_whisper_for_ov_inference, OpenVINOAudioEncoder, OpenVINOTextDecoder


model.encoder = OpenVINOAudioEncoder(core, WHISPER_ENCODER_OV, device=device.value)
model.decoder = OpenVINOTextDecoder(core, WHISPER_DECODER_OV, device=device.value)

Run video transcription pipeline

Now, we are ready to start transcription. We select a video from YouTube that we want to transcribe. Be patient, as downloading the video may take some time.

import ipywidgets as widgets
link = widgets.Text(
    placeholder="Type link for video",

Text(value='', description='Video:', placeholder='Type link for video')
from pytube import YouTube

print(f"Downloading video {link.value} started")

output_file = Path("downloaded_video.mp4")
yt = YouTube(link.value)
print(f"Video saved to {output_file}")
Downloading video started
Video saved to downloaded_video.mp4
from utils import get_audio

audio, duration = get_audio(output_file)

Select the task for the model:

  • transcribe - generate audio transcription in the source language (automatically detected).

  • translate - generate audio transcription with translation to English language.

task = widgets.Select(
    options=["transcribe", "translate"],
    description="Select task:",
Select(description='Select task:', index=1, options=('transcribe', 'translate'), value='translate')
transcription = model.transcribe(audio, task=task.value)

“The results will be saved in the file. SRT is one of the most popular formats for storing subtitles and is compatible with many modern video players. This file can be used to embed transcription into videos during playback or by injecting them directly into video files using ffmpeg.

from utils import prepare_srt

srt_lines = prepare_srt(transcription, filter_duration=duration)
# save transcription
with output_file.with_suffix(".srt").open("w") as f:

Now let us see the results.

widgets.Video.from_file(output_file, loop=False, width=800, height=800)
Video(value=b"x00x00x00x18ftypmp42x00x00x00x00isommp42x00x00:'moovx00x00x00lmvhd...", height='800…
00:00:00,000 --> 00:00:05,000
 What's that?

00:00:05,000 --> 00:00:07,000

00:00:07,000 --> 00:00:10,000
 Hello, humans.

00:00:10,000 --> 00:00:15,000
 Focus on me.

00:00:15,000 --> 00:00:16,000
 Focus on the guard.

00:00:16,000 --> 00:00:20,000
 Don't tell anyone what you've seen in here.

00:00:20,000 --> 00:00:24,000
 Have you seen what's in there?

00:00:24,000 --> 00:00:30,000
 Intel. This is where it all changes.

Interactive demo

import gradio as gr

def transcribe(url, task):
    output_file = Path("downloaded_video.mp4")
    yt = YouTube(url)
    audio, duration = get_audio(output_file)
    transcription = model.transcribe(audio, task=task.lower())
    srt_lines = prepare_srt(transcription, duration)
    with output_file.with_suffix(".srt").open("w") as f:
    return [str(output_file), str(output_file.with_suffix(".srt"))]

demo = gr.Interface(
    [gr.Textbox(label="YouTube URL"), gr.Radio(["Transcribe", "Translate"], value="Transcribe")],
    examples=[["", "Transcribe"]],
except Exception:
    demo.launch(share=True, debug=False)
# if you are launching remotely, specify server_name and server_port
# demo.launch(server_name='your server name', server_port='server port in int')
# Read more in the docs:
Running on local URL:

To create a public link, set share=True in launch().
Keyboard interruption in main thread... closing server.