Updated: Feb 28 2022

In this blog post we describe two case studies to illustrate improvements in speech-to-text or ASR recognition accuracy that can be expected from training of the underlying acoustic models.  We trained our acoustic model to recognize Indian and Irish English better.

Case study setup

The Voicegain out-of-the-box Acoustic Model which is available as default on the Voicegain Platform had been trained to recognize mainly US English although our training data set did contain some British English audio. The training data did not contain Indian and Irish English, except for maybe accidental occurrences.

Both case studies were performed in an identical manner:

• Training data contained about 300 hours of transcribed speech audio.
• Training was done to get improved accuracy on the new type of data but at the same time to also retain the baseline accuracy. An alternative would have been to aim for maximum improvement on the new data at expense of accuracy of the baseline model.
• Training was stopped after significant improvement was achieved. It could have been continued to achieve further improvement, although that might have been marginal.
• Benchmarks presented here were done on data that was not included in the training set.

Case Study 1: Indian English

Here are the parameters of this study.

• We had 250 hours of audio containing male and female speakers, each speaker reading about 50 minutes worth of speech audio.
• We separated 6 speakers for the benchmark, selecting 3 male and 3 female samples. Samples were selected to contain both easy, medium, and difficult test cases.

Here are the results of the benchmark before and after training. For comparison. we also include results from Google Enhanced Speech-to-Text.

‍

‍

Some observations:

• All 6 test speakers show significant improvement over the original accuracy.
• After training the accuracy of 5 speakers is better than Google Enhanced Speech-to-Text. The one remaining speaker improved a lot (from 62% to 76%) but the accuracy was still not as good as Google. We examined the audio and it turns out that it was not recorded properly. The speaker was speaking very quietly and the microphone gain was set very high - this resulted in the audio containing a lot of strange artifacts, like e.g. tongue clicking. The speaker also ready the text in a very unnatural "mechanical" way. Kudos to Google for doing so well on such a bad recording.
• On average custom-trained Voicegain speech-to-text was better by about 2% on our Indian English benchmark compared to Google Enhanced recognizer.

Case Study 2: Irish English

Here are the parameters of this study.

• We collected about 350 hours of transcribed speech audio from one speaker from Northern Ireland.
• For the benchmark we retained some audio from that speaker that was not used for training plus we found audio from 5 other speakers with various types of Irish English accents.  

Here are the results of the benchmark before and after training. We also include results from Google Enhanced Speech-to-Text.

‍

‍

Some observations:

• The speaker that was used for training is labeled here as 'Legge'. We see huge improvement after training from 76.2% to 88.5% which is significantly above Google Enhanced with 83.9%
• The other speaker with over 10% improvement is 'Lucas' which has a very similar accent to 'Legge'.
• We looked in detail at the audio  of the speaker labeled 'Cairns' who had the least improvement and for whom Google was better than our custom trained recognizer. The audio has significantly lower quality that the other samples plus it contains noticeable echo. Its audio characteristics are quite different from that audio characteristics of the training data used.
• On average custom trained Voicegain speech-to-text was better by about 1% on our Irish English benchmark compared to Google Enhanced recognizer.

Further Observations

• The amount of data used in training at 250-350 hours was not large given that normally Acoustic Models for speech recognition are trained on 10s of thousands of hours of audio.
• The large improvement on 'Legge' speaker suggest that if the goal is to improve recognition on very specific type of speech or speaker the training set could be lower, maybe 50 to 100 hours, to achieve significant improvement.  
• Bigger training set may be needed - 500 hours or more - in cases where the variability of speech and other audio characteristics is large.

UPDATE Feb 2022

We have published 2 additional studies showing the benefits of Acoustic Model training:

• Getting high Speech Recognition Accuracy on Alphanumeric Sequences: A Case Study with UK Zip Codes
• Model Training Delivers Huge Gains in Accuracy - Case Study: Indian Food Bot

Interested in Voicegain? Take us for a test drive!

1. Click here for instructions to access our live demo site.

2. If you are building a cool voice app and you are looking to test our APIs, click hereto sign up for a developer account  and receive $50 in free credits

3. If you want to take Voicegain as your own AI Transcription Assistant to meetings, click here.

In our previous post we described how Voicegain is providing grammar-based speech recognition to Twilio Programmable Voice platform via the Twilio Media Stream Feature.

Starting from release 1.16.0 of Voicegain Platform and API it possible to use Voicegain speech-to-text for speech transcription (without grammars) to achieve functionality like using TwiML <Gather>.

The reasons we think it will be attractive to Twilio users are:

• lower cost per each speech-to-text capture
• higher accuracy for customers who choose Acoustic Model customization
• access to all speech-to-text hypotheses in word-tree output mode
  

Using Voicegain as an alternative to <Gather> will have similar steps to using Voicegain for grammar-based recognition - these are listed below.

Initiating Speech Transcription with Voicegain

This is done by invoking Voicegain async transcribe API: /asr/transcribe/async

Below is an example of the payload needed to start a new transcription session:

‍

‍

Some notes about the content of the request:

• we are requesting the callback to return transcript in text form - other options are possible like words (individual words with confidences) and word-tree (words organized in a tree of recognition hypotheses)
• startInputTimers tells ASR to delay start of timers - they will be started later when the question prompt finishes playing
• TWIML is set as the streaming protocol with the format set to PCMU (u-law) and sample rate of 8kHz
• asr settings include the two timeouts used in transcription - no-input, and complete timeouts.

This request, if successful, will return the websocket url in the audio.stream.websocketUrl field. This value will be used in making a TwiML request.

Note, in the transcribe mode DTMF detection is currently not possible. Please let us know if this is something that would be critical to your use case.

TwiML <Connect><Stream> request

After we have initiated a Voicegain ASR session, we can tell Twilio to open Media Streams connection to Voicegain. This is done by means of the following TwiML request:

‍

‍

Some notes about the content of the TwiML request:

• the websocket URL is the one returned from Voicegain /asr/transcribe/async request
• more than one question prompt is supported - they will be played one after another
• three types of prompts are supported: 01) recording retrieved from a URL, 02) TTS prompt (several voices are available), 03) 'clip:' prompt generated   using Voicegain Prompt Manager which supports dynamic concatenation of prerecorded prompts
• bargeIn is enabled - prompt playback will stop as soon as caller starts speaking

Returned Transcription Response

Below is an example response from the transcription  in case where "content" : {"full" : ["transcript"] } .

‍

‍

We want to share a short video showing live transcription in action at CBC. This one is using our baseline Acoustic Model. No customizations were made, no hints used. This video gives an idea of what latency is achievable with real-time transcription.

‍

Automated real-time transcription is a great solution for accommodating hearing impaired if no sign-language interpreter is available. I can be used, e.g., at churches to transcribe sermons, at conventions and meetings to transcribe talks, at educational institutions (schools, universities) to live transcribe lessons and lectures, etc.

‍

Voicegain Platform provides a complete stack to support live transcription:

• Utility for audio capture at source
• Cloud based or On-Prem transcription engine and API
• Web portal for controlling multiple simultaneous live transcriptions
• Web-based viewer app to enable following the transcription on any device with web browser. This app can also be embedded into any web page.

Very high accuracy - above that provided by Google, Amazon, and Microsoft Cloud speech-to-text - can be achieved through Acoustic Model customization.

Voicegain adds grammar-based speech recognition to Twilio Programmable Voice platform via the Twilio Media Stream Feature.

The difference between Voicegain speech recognition and Twilio TwiML <Gather> is:

1. Voicegain supports grammars with semantic tags (GRXML or JSGF) while <Gather> is a large vocabulary recognizer that just returns text, and
2. Voicegain is  significantly cheaper (we will describe the price difference in an upcoming blog post).

When using Voicegain with Twilio, your application logic will need to handle callback requests from both Twilio and Voicegain.

Each recognition will involve two main steps described below:

Initiating Speech Recognition with Voicegain

This is done by invoking Voicegain async recognition API: /asr/recognize/async

Below is an example of the payload needed to start a new recognition session:

‍

‍

Some notes about the content of the request:

• startInputTimers tells ASR to delay start of timers - they will be started later when the question prompt finishes playing
• TWIML is set as the streaming protocol with the format set to PCMU (u-law) and sample rate of 8kHz
• asr settings include the three standard timeouts used in grammar based recognition - no-input, complete, and incomplete timeouts
• grammar is set to GRXML grammar loaded from an external URL

This request, if successful, will return the websocket url in the audio.stream.websocketUrl field. This value will be used in making a TwiML request.

Note, if the grammar is specified to recognize DTMF, the Voicegain recognizer will recognize DTMF signals included in the audio sent from Twilio Platform.

TwiML <Connect><Stream> request

After we have initiated a Voicegain ASR session, we can tell Twilio to open Media Streams connection to Voicegain. This is done by means of the following TwiML request:

‍

‍

Some notes about the content of the TwiML request:

• the websocket URL is the one returned from Voicegain /asr/recognize/async request
• more than one question prompt is supported - they will be played one after another
• three types of prompts are supported: 01) recording retrieved from a URL, 02) TTS prompt (several voices are available), 03) 'clip:' prompt generated   using Voicegain Prompt Manager which supports dynamic concatenation of prerecorded prompts
• bargeIn is enabled - prompt playback will stop as soon as caller starts speaking

Returned Recognition Response

Below is an example response from the recognition. This response is from built-in phone grammar.

‍

‍

Some of the feedback that we received regarding the previously published benchmark data, see here and here, was concerning the fact that the Jason Kincaid data set contained some audio that produced terrible WER across all recognizers and in practice no one would user automated speech recognition on such files. That is true. In our opinion, there are very few use cases where WER worse than 20%, i.e. where on average 1 in every 5 words is recognized incorrectly, is acceptable.

New Methodology

What we have done for this blog post is we have removed from the reported set those benchmark files for which none on the recognizers tested could deliver WER 20% or less. This criterion resulted in removal of 10 files - 9 from the Jason Kincaid set of 44 and 1 file from the rev.ai set of 20. The files removed fall into 3 categories:

• recordings of meetings - 4 files (this amounts to half of the meeting recordings in the original set),
• telephone conversations - 4 files (4 out of 11 phone phone conversations in the original set),
• multi-presenter, very animated podcasts - 2 files (there were a lot of other podcasts in the set that did meet the cut off).

The results

As you can see, Voicegain and Amazon recognizers are very evenly matched with average WER differing only by 0.02%, the same holds for Google Enhanced and Microsoft recognizer with the WER difference being only 0.04%. The WER of Google Standard is about twice of the other recognizers.

‍

[UPDATE - October 31st, 2021:  Current benchmark results from end October 2021 are available here. In the most recent benchmark Voicegain performs better than Google Enhanced. Our pricing is now 0.95 cents/minute]

[UPDATE: For results reported using slightly different methodology see our new blog post.]

This is a continuation of the blog post from June where we reported the previous speech-to-text accuracy results. We encourage you to read it first, as it sets up a  context to better understand the significance of benchmarking for speech-to-text.

Apart for that background intro, the key differences from the previous post are:

• We have improved our recognizer and we are now essentially tied with Amazon
• We added another set of benchmark files - 20 files published by rev.ai. Please reference the data linked here when trying to reproduce this benchmark.

Here are the results.

Comparison to the June benchmark on 44 files.

‍

Less than 3 months have passed from the previous test, so it is not surprising to see no improvement on Google and Amazon recognizers.

Voicegain recognizer has how overtaken Amazon by a hair breadth in average accuracy, although Amazon median accuracy on this data set is slightly above Voicegain.

Microsrosoft recognizer has improved during this time period - on the 44 benchmark files it is now on average better than Google Enhanced (in the chart we retained ordering from the June test). The single bad outlier in Google Enhanced results does alone not account for the better average WER on the Microsoft on this data set.  

Google Standard is still very bad and we will likely stop reporting on it in detail in our future comparisons.

Results from the benchmark on 20 new files.

The audio from the 20-file rev.ai test is not as challenging as some of the files in the 44-file benchmark set. Consequently the results are on average better but the ranking of the recognizers does not change.

As you can see in this chart, on this data set the Voicegain recognizer is marginally better than Amazon in.  It has lower WER on 13 out of 20 test files and it beats Amazon in the mean and median values. On this data set Google Enhanced beats Microsoft.

Combined results on 44+20 files

Finally, here are the combined results for all the 64 benchmark files we tested.

‍

On the combined benchmark Voicegain beats Amazon both in average and median WER, although the median advantage is not as big as on the 20 file rev.ai set. [Note that as of 2/10/21 Voicegain WER is now 16.46|14.26]

What we would like to point out is that when comparing Google Enhanced to Microsoft, one wins if we compare the average WER while the other has a better median WER value. This highlights that the results vary a lot depending on what specific audio file is being compared.

Conclusions

These results show that choosing the best recognizer for a given application should be done only after thorough testing. Performance of the recognizers varies a lot depending on the audio data and acoustic environment. Moreover, the prices vary significantly. We encourage you to  try the Voicegain Speech-to-Text engine for your application. It might be a better fit for your application. Even if the accuracy is a couple of points behind the two top players,  you might still want to consider Voicegain because:

• Our acoustic models can be customized to your specific speech audio and this can reduce the word error rates below the best out-of-the-box options - see our Improved Accuracy from Acoustic Model Training blog post.
• If the accuracy difference is small, Voicegain might still make sense given the lower price.  
• We are continuously training our recognizer and it is only a matter of time before we catch up.

‍