How Awaaz listens, and where it falls short.

This page explains the system in plain language, with enough detail for technical and professional scrutiny. It describes what is deployed today, not what we hope to build.

What the system is for

Awaaz checks whether a specific target speech sound was recognised in a recording of a single word or short phrase, and, when it is clear, which sound was recognised instead.

Intended use

Encouraging feedback during home practice.
A structured overview of supported sounds after a Speech Check.
Background information a person may choose to share with a professional.

Not intended for

Diagnosing any speech, language or hearing condition.
Making decisions about someone’s care, education or eligibility for services.
Assessing sounds, voice, fluency or language outside the supported list.

Methodology

General speech-to-text is the wrong tool: it is built to guess the intended word, so it quietly “corrects” the very differences practice is about. Awaaz works at the level of individual speech sounds instead.

1
Capture
The browser records a short clip with noise suppression, echo cancellation and automatic gain turned off, because those filters erase short consonant bursts. The microphone opens about 200 ms before the prompt and stays open about 350 ms after you tap stop, so the start and end of the word are kept. A loudness check against the room’s own noise floor rejects silence before anything is scored.
2
Recognise
A wav2vec2 phoneme recogniser (facebook/wav2vec2-lv-60-espeak-cv-ft) turns the audio into a sequence of phonetic symbols. There is no word-level language model, so it cannot replace what was said with the word it expected.
3
Target
The target word is converted to the same symbols with espeak-ng (English or Urdu). Where espeak gets a word wrong, the lesson carries an explicit target, written and tested by hand.
4
Align and score
The two sequences are aligned (Needleman–Wunsch), giving each target sound as correct, substituted or omitted. The lesson’s focus sound is weighted and gates the result, so a substitution of that sound is never outweighed by the rest of the word being right.
5
Account for known artefacts
Some differences come from the recogniser, not the speaker. Small equivalence classes absorb these, scoped per language because each set measurably broke scoring in the other. They never absorb the substitutions practice looks for, such as W for R or S for TH.

The scoring service is stateless: it decodes the audio in memory, returns the result and keeps nothing.

Data

Awaaz has not collected speech recordings to train or tune any model, and does not use people’s recordings for research.

The recogniser is a published model. According to its publishers it was pre-trained on unlabelled English audiobook speech and fine-tuned on multilingual Common Voice recordings to output phonetic symbols. It was not trained on children’s speech or on speech differences, and Awaaz has not re-trained it.

We do not yet have demographic or accent coverage figures for the recogniser’s training data that we can verify, so we don’t publish any.

Evaluation so far

Awaaz has not yet published a formal evaluation comparing its observations with assessments by speech-language therapists, and has not measured accuracy on children’s speech at scale. Until it has, no accuracy figure is claimed and results should be read as practice feedback.

During development, lesson words were tested with real and, in some cases, synthesised recordings. Words the system could not score fairly were removed or adjusted rather than kept.

Examples from development testing

English “ship”: The final P was recognised as V, P, B, K or nothing across six attempts. Retired in favour of “fish”.
Urdu ق: Recognised as ک almost every time. Retired from the lessons rather than merged.
Urdu مچھلی: Aspiration results tracked speaking speed rather than pronunciation. Replaced with بادشاہ.
Urdu گائے: espeak’s literary reading scored natural speech at 50%. An explicit everyday target now scores the natural form at 100% and the literary form at 67%.
Urdu ت: Correct attempts were all labelled د. Merged with د, which means a د substitution now passes. A known loss, documented on the profile card.

Professional involvement

Professional contributors will be named here, with their permission and a description of their actual role. Learn & Explore articles show whether a speech-language therapist has reviewed them; at the moment none has.

Known limitations

These apply to every result Awaaz shows.

The recogniser was trained on adult speech. Children’s speech and speech differences score less reliably.
Background noise, distance from the microphone and low-quality microphones change what is recognised.
Close vowels (“sock” and “suck”) and “C” for “S” return identical results, so vowels are not assessed.
Urdu ت cannot be told apart from د.
Urdu ق, aspirated sounds such as چھ, and Urdu vowels are not scored.
Regional accents of Urdu and English have not been evaluated separately.
A check is a handful of recordings at one moment. It does not show how someone speaks in conversation.
Results from two checks are only compared when the same prompts and language were used, and even then differences may reflect recording conditions.

Privacy and security

Recordings travel over an encrypted connection to the scoring service and are not stored there. Profiles, history and any recordings you save stay in your browser on your device, with deletion and export under your control.

Read the privacy summary

Updates

October 2026New website and app. Awaaz Speech Check with recording-quality checks and per-sound observations. Optional on-device recordings with automatic deletion. Therapist plan links.
EarlierLessons for English “ship”, Urdu ق, مچھلی and شترمرغ retired after testing. Explicit targets added for گائے, آیا and isolated sounds.