An interactive Streamlit app that analyses recorded or uploaded speech for vocal fluency, filler words, long pauses, and acoustic clarity — using Whisper for transcription, Praat-Parselmouth for acoustic feature extraction, and custom heuristics for fluency scoring.

Built as a tool for language learners and public speakers who want objective, data-driven feedback on their delivery.

The app is designed around two layers of analysis:

• Surface layer (available in the live demo): Whisper transcription, filler word detection, pause flagging, and Praat-based acoustic metrics (pitch, intensity, jitter, shimmer).

• Phoneme layer (local only): When Montreal Forced Aligner (MFA) is available, the app additionally aligns the transcript to the audio at the level of individual sounds — enabling precise timing analysis of specific phonemes. This opens the door to pronunciation feedback at a granularity that word-level transcription alone cannot achieve.

Live demo: https://speech-analyser.streamlit.app/

⚠️ The live demo runs the full Whisper + Praat pipeline. MFA phoneme alignment requires local setup and is disabled on the hosted version.

• Upload or record your own voice
• Detect filler words (customisable list)
• Flag long pauses above a configurable threshold
• Analyse pitch, intensity, jitter and shimmer via Praat
• Visual feedback with coaching-style summaries
• Phoneme-level alignment via MFA (local only — see below)

Acoustic Analysis — Intensity over time

Acoustic Analysis — Pitch and Jitter over time

git clone https://github.com/alexdimmock95/speech-analyser.git
cd speech-analyser

python3 -m venv venv
source venv/bin/activate  # Windows: venv\Scripts\activate

pip install -r requirements.txt

streamlit run app.py

• streamlit — web app interface
• numpy — numerical operations
• matplotlib — plotting
• librosa — audio processing
• soundfile — audio file I/O
• parselmouth — Praat integration for acoustic analysis
• openai-whisper — speech transcription

• FFmpeg — required by librosa and Whisper:
  • macOS: brew install ffmpeg
  • Ubuntu: sudo apt-get install ffmpeg
  • Windows: download and add to PATH

Montreal Forced Aligner (MFA) adds phoneme-level timing to the analysis — it aligns the transcript to the audio at the level of individual sounds rather than words. This enables more precise fluency analysis but requires a separate local environment due to its dependencies.

MFA has heavy dependencies that conflict with the main app packages. The app calls MFA via a subprocess, keeping your working environment clean.

conda create -n mfa_env python=3.10
conda activate mfa_env
pip install montreal-forced-aligner

You will also need to download the English acoustic model and pronunciation dictionary from the MFA model repository.

⚠️ The acoustic model and dictionary files are too large to include in this repo (~2GB+). Download them separately and update the paths in run_mfa.sh to point to your local copies.

The MFA component uses run_mfa.sh. On Windows, use WSL or Git Bash.

├── app.py              # Main Streamlit app
├── run_mfa.sh          # Bash script for MFA alignment
├── requirements.txt    # Python dependencies
├── output/             # Analysis outputs (not tracked in Git)

Jitter and shimmer measurements are sensitive to background noise, microphone quality, and recording environment. In noisy conditions, Praat may produce elevated readings that don't reflect genuine vocal instability. Similarly, pitch tracking can misfire on unvoiced segments or noise, producing outlier Hz values that skew statistics.

For best results, use a clean microphone input in a quiet environment.

The current feedback labels (e.g. 🔴 High jitter, 🔵 Low vocal energy) are applied globally across the entire recording. This means a speaker who is confident for most of a recording but hesitates briefly in one section will receive the same label as someone who hesitates throughout. Richer feedback would require time-localised labelling — flagging where in the audio a metric deteriorates, not just whether it does overall.

Montreal Forced Aligner is powerful but difficult to deploy in cloud environments due to its size (~2GB+ models) and conda-based dependency structure. The phoneme alignment component is therefore only available when running the app locally with MFA installed in a separate conda environment. A future version might explore lighter forced alignment alternatives such as wav2vec2-based alignment, which can run without MFA entirely.

• Time-localised feedback — flag specific moments in the audio where pitch, intensity or fluency metrics deteriorate, rather than summarising across the whole recording
• Richer coaching labels — move from binary labels (high/low) to contextual feedback tailored to the speaker's pattern across the recording
• Lightweight phoneme alignment — explore wav2vec2-based forced alignment as a cloud-deployable alternative to MFA
• Multi-language support — extend transcription and analysis to non-English input, leveraging Whisper's multilingual capabilities
• Visual phoneme timeline — display phoneme boundaries overlaid on the acoustic plots

• Montreal Forced Aligner
• Praat-Parselmouth

MIT License — see LICENSE for details.