Kaldi

What do you like best about Kaldi?

The upsides of Kaldi is that once you know it very deeply after a lot of experience, the possibilities become quite endless for customising acoustic models. The user community for Kaldi is quite vast, interactive, and odds are that someone has had the same problem as you if you just know what to look for. There are many useful tools in the utils/ folder, even though they all need thorough customisation for appropriate use for the model building, as the process is inherently data-driven. Kaldi does feel like a massive puzzle, and piecing it together is quite rewarding in a strange, masochistic way. It's great that since it is community-based, there are many pre-existing recipes that are easily customisable for various use cases and that you can contribute with your own recipe. My own holy grail that I always go back to is the Eleanor Chodroff tutorial for building Kaldi acoustic models, since it describes the particular data structure required for the process.

What do you dislike about Kaldi?

Well. There are many issues that i must adress pertaining to Kaldi. This is just some of those things that everyone knows and has accepted, but bottom line is that currently Kaldi is not user friendly or intuitive. While there are a lot of recipes, they are all border-line useless because they all need to be thoroughly customised as the point of creating a custom ASR model is that it is entirely data-driven. There are no explanations as to what the many utilities are or why they must occur in which order. The only way to learn how to use Kaldi is through thorough trial and error. If you try to ask Dan Povey questions on the forum, you will get a passive-agressive response thinly veiled as advice telling you to switch careers and stop doing speech recognition. The entire framework is so un-intuitive that it maketh no sense. Literally any user interface or some more comprehensive and straight forward instruction would be great.

What also annoys me is that there are so many fantastic language representation systems with which one can make a great LM, but since Kaldi only works with ARPA format, it disallows any great progress in the quality of ASR in regards to LMs.

Another thing is that if you make one mistake, you pretty much have to start all over again.

Especially since Kaldi is so data-driven, it is particularly difficult to automate AM building processes which is hindering to company growth if Kaldi is the main tool that is used there.

Recommendations to others considering Kaldi:

Patience, trial and error, and lots of reading is key. It can be a powerful tool and create great models, but starting out can feel quite slow.

What problems is Kaldi solving and how is that benefiting you?

I use Kaldi to build custom ASR models for clients, the benefits were that it was pretty much the only sensible option for ASR building at the time.

What do you like best about Kaldi?

Language Model creation and FST creation.

What do you dislike about Kaldi?

Lexicon generation requires linguists help if open source lexicon data is not available.

Recommendations to others considering Kaldi:

If someone want to understand the practical experience of ASR working.

What problems is Kaldi solving and how is that benefiting you?

Domain specific ASR.

What do you like best about Kaldi?

Speed, accuracy. It makes the job simpler. Speed was great. All the documentation was there. The instruction was really helpful. There is no other tool like kaldi to implement the speech-to-text conversion.

What do you dislike about Kaldi?

Operating system compatibility. I faced a problem with windows OS. Kaldi was faster in Linux but it was difficult to implement in windows.

Recommendations to others considering Kaldi:

Highly recommend kaldi for speech text conversion. Easy to use and implement. Instruction for implementation was well documented and each step was very clear.

What problems is Kaldi solving and how is that benefiting you?

Speech to text conversion. understanding the speech data using kaldi was awesome.

What do you like best about Kaldi?

It has fst for LM which makes it very flexible and customizable solution to target application domain. It also renders the phoneme time stamps in ctm output, which makes it an ideal solution for time synchronization and confidence score calibration

What do you dislike about Kaldi?

It needs a lots and lots of memory resources to load the bulky acoustic models and the LM graphs.

What problems is Kaldi solving and how is that benefiting you?

Fluency and Pronunciation scoring for kids edutech. We could precisely score the confidence scores for spoken sentence based from deterministic lattice indexing and precise time synchronization.