ASR

Practical considerations for developers considering OpenAI's Whisper ASR

On March 1st, Open AI announced that developers could now access the Whisper Speech-to-Text model via easy to use APIs. OpenAI also released APIs to GPT3.5, the LLM behind the buzzy ChatGPT product.

Since Whisper's initial release in October 2022, it has been a big draw for developers. A highly accurate Open Source ASR is extremely compelling. Whisper has been trained on 680,000 hours of audio data which is much more than what most models are trained on. Here is a link to their github.

However there were two major limitations. 1) Running Whisper requires expensive memory-intensive GPU based compute options (see below). 2) A company still had to invest in an engineering team that could test, run and support the model in a production environment.

By taking over the responsibility of hosting this model and making it accessible via easy-to-use APIs, Open AI addresses both of the above limitations.


This article highlights some of the key strengths and limitations of using Whisper - whether using Open AI's APIs or hosting it on your own.


Strengths

1. Accuracy

In our benchmark tests, Whisper models demonstrated high accuracy for a widely diverse range of audio datasets. Our ML engineers found that the Whisper models perform well on a wide range of audio datasets - from meetings, classroom lectures, YouTube videos and call center audio. We benchmarked Whisper-base, Whisper-small and Whisper-medium.

The median Word Error Rate (WER) for Whisper-medium was 12.5% for meeting audio and 17.5% for call center audio. This was indeed better than the WERs of other large players like AWS, Azure and Google. However here is an interesting fact - it is possible to match and even exceed Whisper's accuracy with custom models. Custom models are models that are trained on our client's data.

Please contact us via email (support@voicegain.ai) if you would like to review these accuracy benchmarks.

2. Affordable, but with a caveat

Whisper's pricing at $0.006/min is much lower than the Speech-to-Text offerings of some of the other larger cloud players. This translates to a 75% discount to Google Speech-to-Text and AWS Transcribe (based on pricing as of the date of this post). However there are a few caveats to this pricing which are outlined in the Limitations section below

3. Whisper API + ChatGPT API, built to be used together

What was also significant was Open AI announced the release of ChatGPT APIs with the release of Whisper APIs. For developers building Voice AI apps, they can now can combine the power of Whisper Speech-to-Text models with the GPT 3.5 LLM (the underlying models that ChatGPT APIs give access to) and they can build really cool apps - whether it is for Meetings or Call Center.

Limitations

1. Does not support Streaming/Real-time

Whisper currently does not support apps that require real-time/streaming transcription which are relevant both to call center and the meetings use-case. While there are some hacks and work-arounds, they are not practical for a production deployment.

2. Expensive to run Whisper on Edge (Datacenter or Private Cloud) 

The throughput of Whisper models - both for the small and medium models - is quite low. Our ML engineers tested Whisper models on popular NVIDIA GPU-based compute instances available in public clouds (AWS, GCP and Microsoft Azure). Net-net, we determined that while developers would not have to pay for software licensing, the cloud infrastructure costs would be so high. We determined that the infrastructure cost of running Whisper - so that it can perform well is in the range of $0.10 - $0.15/hour.  

In addition to this infrastructure cost the larger expense of running Whisper on the Edge (On-Premise + Private Cloud) is that it would require a dedicated back-end Engineering & Devops team to run the model in a cost-effective manner.

b) Price per channel makes it expensive for Call Center & Meeting use-case

As of the publication of this post, Whisper does not have a multi-channel audio API. So if your application involves audio with multiple speakers, then Whisper's effective price-per-min = Number of channels * 0.006. For both meetings and call center use-cases, this pricing can become prohibitive. 

3. Missing Key Features  - Diarization, Time-Stamps, PII Redaction

This release of Whisper is missing some key features that developers would need. The three important features we noticed are Diarization (speaker separation), Time-stamps and PII Redaction. 

About Voicegain

At Voicegain, we have built deep-learning-based Speech-to-Text/ASR models that match the accuracy of models from the large players. For over 3 years nows, developers have been using our Speech-to-Text APIs to build and launch successful products. Our focus has been on voice developers that need high accuracy (achieved by training custom acoustic models) and deployment in private infrastructure at an affordable price. We provide an accuracy SLA where we guarantee that a custom model that is trained on your data will be as accurate if not more than most popular options including Open AI's Whisper. Here is a link to our

We also have models that are trained specifically on call center audio - so if you are looking for a call center focused model we can provide higher accuracy than Whisper.

While Whisper is a worthy competitor (of course a much larger company with 100x our resources), as developers we welcome the innovation that Open AI is unleashing in this market. By adding ChatGPT APIs to our Speech-to-Text , we are planning to broaden our API offerings to developer community.

To sign up for a developer account on Voicegain with free credits, click here.

Sign up for an app today
* No credit card required.

Enterprise

Interested in customizing the ASR or deploying Voicegain on your infrastructure?

Contact Us → 
Voicegain - Speech-to-Text
Under Your Control