Module 1The speech problem and its cost
3 lessonsThe speech problem and its cost. Price the mistakes on 4,000 calls a month, see sound as numbers and why a phone line loses sounds, and set up the speech project with one call and its reference.
Turn phone calls into text and decisions, and build a phone line that answers by itself. Measure speech recognition honestly on phone-quality audio, compare voice detectors, recognisers, speaker splitting and voices on the same calls, build a small IVR that understands speech and keypad tones, and compare it with a voice agent.
Read every lesson as a guest, or sign in to enroll and track your progress.
Enroll to save the lessons that you complete. It is free.
You are enrolled. From now on, the lessons that you complete are saved.
This course teaches speech as a business system: audio as data, honest measurement of speech recognition, voice activity detection, local recognisers compared by slice, speaker splitting and timing, call reasons, summaries and compliance checks from transcripts, text-to-speech, a small IVR with speech and keypad tones that learners build and test, a voice agent with a latency budget, and privacy and misuse risks of voices. It does not teach how neural networks and transformers learn, prompt design from the start, or the platform that deploys and monitors models.
Take another course if you want to:
By the end of this course, you will be able to:
Larkfield’s help desk takes about 4,000 phone calls a month in English and French. The calls are recorded with consent, but nobody can search them. Nobody knows why most people call, which promises the agents made, or which calls need a follow-up. At night, the phone simply rings out.
In this course, you follow Priya, Larkfield’s data scientist, Grace, the support lead, and Malik, the telephony engineer, as they turn calls into decisions. You start by measuring: you build a test set, write a normalisation guide, and count the errors that matter, such as a wrong order number, apart from the overall score. You test every method the same way, on the same calls, in one comparison table. In the command-spotting case study, you see a small, specialised model beat a large one where it counts.
Then you go further: voice detectors that find the speech and stop a recogniser from inventing text in silence, local recognisers of different sizes on a laptop CPU, speaker splitting, call reasons and summaries from transcripts, and compliance checks.
You build one project through the course: the Larkfield IVR, a small phone line that you call from a test phone in your browser. It understands speech and keypad tones, reads back the order number it heard, asks again after a miss and hands over to a person after two. You test it with recorded calls, then compare it with a voice agent and a hybrid on the same calls. You finish with the production work: cost per audio hour, card numbers removed from text and audio, consent and retention, and why a voice is not a password. Every Larkfield call is synthetic, made with open text-to-speech voices; no real customer call is used.
You completed Machine Learning, From Problem to Reliable Model, or you have the same skills. You can read and change Python code, use notebooks and a terminal, and read JSON, and you know test sets, precision and recall, thresholds and slices. The deep learning and LLM API courses help but are not required. You need Python 3.12 on your computer, or a free Google or Kaggle account for the notebooks, and headphones. You need no key, no graphics card and no money.
Courses to take first: Machine Learning: From Problem to Reliable Model.
AI and ML engineers, backend developers and data scientists who must turn speech into text and decisions, such as call centres, meetings, voice notes and captions, or build a phone line or voice assistant, and technical leads who must choose between local and hosted speech models.
Experienced AI practitioners designed this course. AI helped us draft the content; people chose the teaching approach and the examples, and reviewed every lesson. The outputs that you see are the actual results of running the code. How we make our courses
The speech problem and its cost. Price the mistakes on 4,000 calls a month, see sound as numbers and why a phone line loses sounds, and set up the speech project with one call and its reference.
Measure before you choose. Word and character error rate step by step, a written normalisation guide, and the errors that matter most, order numbers, product names and negations, counted apart by slice.
Find the speech. An energy threshold, WebRTC VAD and Silero VAD compared on music on hold and background talk, and how cutting silence stops invented text and saves cost.
Speech recognition, compare the families. Vosk and Whisper sizes on a laptop CPU by error, speed and memory, hosted services, the Larkfield vocabulary, and two languages on noisy lines.
Who spoke when. Speakers split by channel, speaker embeddings and clustering on mono calls, pretrained pipelines and their terms, and word times that say who said what and when.
From transcript to decision. Call reasons by rules, a classifier and an LLM, structured summaries checked against the call, and compliance phrases found in the transcript and in the audio.
Build a small IVR. A phone menu as a map of states, voices for the prompts, keypad tones, speech with a keypad fallback, order numbers read back and confirmed, a handover after two misses, and recorded test calls.
From IVR to voice agent. What an LLM adds and costs, a latency budget for one turn, turn-taking and interruptions, evaluating a voice agent as a whole, and the IVR, the agent and a hybrid on the same calls.
Run it in production, safely. Batch and streaming, cost per audio hour, card numbers removed from text and audio, consent and retention, voice cloning and speaker checks, and watching the system over time.
Four end-to-end case studies on real open speech. Recognisers for two languages, meeting transcripts with speakers, command spotting and an IVR for spoken bank requests. Complete at least one for the certificate; each one is a portfolio piece.
Complete at least one case study for the certificate. The others are optional. What each one covers
Use every module in one practice project. You collect evidence, and check it against the course outcomes.
Realistic projects from start to finish, on real data, with every output explained. Each one makes a portfolio piece. Complete at least one for the certificate; the others are optional.
Choose a recogniser for two languages2 hr
Compare Vosk and Whisper sizes on real English and French speech from an open dataset by speaker group, and recommend one with error, speed and cost.
Meeting transcripts with speakers2 hr 30 min
Diarize and transcribe real recorded meetings from an open corpus, measure DER and WER, and make action-item notes.
Spot commands on a small device2 hr 30 min
Build a keyword spotter on an open spoken-commands set, compare a small CNN on spectrograms with Whisper tiny, and measure false alarms per hour.
An IVR for spoken intents2 hr 30 min
Build a menu-and-speech IVR for a bank's common requests, test it with open spoken-intent recordings by speaker group, and compare keyword rules, a classifier and an LLM for the intent step.
Loading the reviews…
Took this course? Sign in to review it. Learners can write a review after they complete 25% of the lessons.