Whisper, released by OpenAI in late 2022, till date has a near-SoTA performance across English & Multi-lingual benchmarks.

The model was trained to do two key speech recognition tasks:

1. Transcribe a given audio in its base language. i.e. take the audio in language "X" and transcribe it.
2. Directly translate an audio to English. i.e. take audio in language "X" and transcribe into English.

As the world grows more and more connected, the need for high quality content is ever-so-increasing. One of the ways to make content more accessible (specially audio), is by transcribing it into different languages, thereby ensuring that the knowledge is spread. ⚡️

The typical workflow for transcribing from audio in language "X" to another language is as follows:

1. Translate and transcribe the audio in language "X" to English. (Base Whisper behaviour)
2. Translate the transcriptions from language "X" to another language. (Typically done with a LLM, you can use for example GPT-3.5/4)

This works great! However, as with any other process, the more steps you run, the higher the chances for error creep.

TL;DR - Yes! Keep in mind that this is a hack, but it seems to work pretty well in our tests! These notes describe how to do it, but serious use of the technique would have to be validated much more throughly! This is because the model wasn't trained on the task we'll use it for, so results may not be as reliable.

Alright, let's get to it! To demonstrate how this works, let's try to transcribe an audio in english (en) language to german (de), italian (it), spanish (es), dutch (nl) and french (fr).

For a more interactive experience you can follow along with this colab!

Note: This tutorial assumes that you have run huggingface-cli login or used notebook_login() to authenticate with the hub. We only need this to access Common Voice, you can safely ignore this step if you run inference on your own audio files or public datasets.

!pip -q install transformers datasets huggingface_hub

Let's instantiate our speech recognition pipeline! For the purpose of colab demo, we'll use a Whisper-large-v2 checkpoint in half-precision (fp16). If you have access to a larger GPU VRAM then remove the torch_dtype arg 🤗

import torch
from transformers import pipeline

whisper_asr = pipeline(
    "automatic-speech-recognition", 
    model="openai/whisper-large-v2",
    torch_dtype=torch.float16,
    device="cuda:0"
    )

To keep things simple, we'll use the Common Voice dataset from the 🤗 Hub in streaming mode & resample the audio to 16KHz as expected by Whisper.

from datasets import load_dataset
from datasets import Audio

common_voice_en = load_dataset("mozilla-foundation/common_voice_11_0", "en",
                               revision="streaming",
                               split="test",
                               streaming=True,
                               use_auth_token=True)

common_voice_en = common_voice_en.cast_column("audio",
                                              Audio(sampling_rate=16000))

Since we cannot render audio in markdown, let's take a look at the transcription.

next(iter(common_voice_en))["sentence"]

output:

Reading metadata...: 16354it [00:00, 31433.60it/s]
'Joe Keaton disapproved of films, and Buster also had reservations about the medium.'

Let's create a wee list of languages to transcribe too.

list_of_languages = ["de", "it", "es", "nl", "fr"]

We essentially force Whisper to decode in one specific language. Because Whisper was trained on 600K+ hours of data it is able to do so fairly well.

So the only change to make this happen would be to set the task as transcribe and change the target language.

for lang in list_of_languages:
    whisper_asr.model.config.forced_decoder_ids = (
        whisper_asr.tokenizer.get_decoder_prompt_ids(
            language=lang,
            task="transcribe"
            )
        )
    print(whisper_asr(next(iter(common_voice_en))["audio"]["array"])["text"])

output:

Reading metadata...: 16354it [00:00, 33718.24it/s]
 Joe Keaton hat Filme verabschiedet und Buster hatte auch Reservations über die Medien.
Reading metadata...: 16354it [00:00, 27172.67it/s]
 Joe Keaton ha disapprovato i film e Buster ha anche delle riservazioni sui media.
Reading metadata...: 16354it [00:00, 41110.13it/s]
 Joe Keaton disaproveció de los filmes y Buster también tenía reservaciones sobre el medio.
Reading metadata...: 16354it [00:00, 39696.06it/s]
 Joe Keaton onverstaanbaar van de films en Buster had ook bewaarschuwingen over de media.
Reading metadata...: 16354it [00:00, 39813.59it/s]
 Joe Keaton a dénoncé les films et Buster avait des réservations sur le médium.

Voila! it works! We successfully transcribed an english audio to other languages.

Some of these translations are a bit off, but we can fix them with some neat generation techniques like constrastive search!

You can use contrastive search by providing penalty_alpha and top_p to the generate_kwargs in the pipeline. You can read more about it here. 🤗

for lang in list_of_languages:
    whisper_asr.model.config.forced_decoder_ids = (
        whisper_asr.tokenizer.get_decoder_prompt_ids(
            language=lang,
            task="transcribe"
            )
        )
    print(whisper_asr(
        next(iter(common_voice_en))["audio"]["array"], 
        generate_kwargs = 
         {
              "penalty_alpha": 0.6, 
              "top_k": 5,
         }
        )["text"])

output:

Reading metadata...: 16354it [00:00, 39409.04it/s]
 Joe Keaton verabschiedete sich von Filmen und Buster hatte Regeltäusen über die Medien.
Reading metadata...: 16354it [00:00, 34203.76it/s]
 Joe Keaton disapprovò i film e Buster aveva anche riservazioni sui media.
Reading metadata...: 16354it [00:00, 24372.39it/s]
 Joe Keaton aprovechó de los filmes y Buster también tenía reservaciones sobre el medio.
Reading metadata...: 16354it [00:00, 41170.46it/s]
 Joe Keaton onvoldoende films en Buster had ook besluitingen over het medium.
Reading metadata...: 16354it [00:00, 23721.35it/s]
 Joe Keaton n'approuve pas les films et Buster avait également des préjugés sur le media.

Notice the subtle differences in the transcription, and how it still gets some things wrong here and there. For your actual use-case, I'd recommend tuning these parameters a bit or use one of the fine-tuned models on the hub.

Good luck! 🤝

1. Contrastive Search often results in over-generation, find strategies to reduce the over-generation.
2. Run a benchmark on FLoRES dataset.
3. Test the benchmark for fine-tuned Whisper models.

Help is more than welcome! Just open an issue or PR and we can work together on this! 🤗