speech-analyser

Transcribes audio and video files and returns speech metrics: word count, speaking rate, filler word detection, silence ratio, and a quality score with natural-language insights. Optionally identifies individual speakers.

Part of the analyser family.

Install

pip install speech-analyser

Requires Python 3.11+. Uses CPU by default; GPU (CUDA) is used automatically if available.

For speaker diarization, install the extra and set a Hugging Face token:

pip install "speech-analyser[diarization]"
export HF_TOKEN=hf_...

Usage

Python

from speech_analyser import SpeechAnalyser

lens = SpeechAnalyser()             # model_size="base" by default
result = lens.analyse("recording.mp3")

m = result["speech_metrics"]
print(f"Duration:  {result['duration']:.1f}s")
print(f"Words:     {m['word_count']} ({m['speaking_rate_wpm']} wpm, {m['pace_category']})")
print(f"Quality:   {m['quality_score']}/100")
print(result["transcript"])

CLI

# Human-readable summary
speech-analyser recording.mp3

# Larger model for better accuracy
speech-analyser lecture.wav --model small

# Machine-readable JSON
speech-analyser recording.m4a --json

# Speaker diarization
speech-analyser interview.mp3 --diarize

# Video input, transcript file + SRT subtitles, timestamped lines
speech-analyser lecture.mp4 -o lecture.txt --srt --timestamps

# Force language, skip silence with the VAD filter
speech-analyser podcast.mp3 --language en --vad

# Start the HTTP server
speech-analyser serve --port 8001

File outputs (-o, --srt, --timestamps) were ported from a standalone transcribe.py script. -o writes the transcript to a text file; --srt writes SubRip subtitles next to the -o file (or next to the input); --timestamps prefixes each line with [HH:MM:SS.mmm]. Confirmations go to stderr, so --json output on stdout stays machine-readable.

HTTP API

curl -X POST http://localhost:8001/analyse \
  -F "file=@recording.mp3"

Supported formats

Audio: .mp3 .wav .m4a .ogg .flac .aac .wma .opus

Video: .mp4 .m4v .mov .avi .mkv .webm — the audio track is extracted to a temporary 16 kHz mono WAV with ffmpeg (must be on PATH; only needed for video inputs).

Model sizes

Models download on first use (~75 MB for base, ~1.5 GB for large-v3).

Output

{
  "transcript": "Good morning everyone...",
  "language": "en",
  "duration": 62.4,
  "segments": [{"start": 0.0, "end": 3.2, "text": "Good morning everyone", "speaker": null}],
  "speech_metrics": {
    "word_count": 120,
    "speaking_rate_wpm": 115.4,
    "pace_category": "natural",
    "filler_word_count": 3,
    "filler_word_rate": 0.025,
    "filler_words_found": ["um", "basically"],
    "silence_ratio": 0.18,
    "actual_speaking_time": 51.2,
    "quality_score": 78,
    "quality_factors": {"clarity": 23, "depth": 18, "balance": 18, "pace": 19},
    "quality_ratings": {"clarity": "excellent", "depth": "good", "balance": "good", "pace": "good"},
    "insights": {
      "strengths": ["Very few filler words — speech is clear"],
      "observations": ["Speaking rate is slightly slow — aim for 130–170 wpm"]
    }
  },
  "diarization_available": false,
  "speakers": null,
  "talk_time": null,
  "file_path": "/path/to/recording.mp3",
  "file_size": 2048000
}

When diarization is enabled, speakers contains per-speaker word count, duration, and percentage; talk_time.is_balanced flags whether one speaker dominates.

The analyser family

Low-level analysis tools. Each accepts files directly and returns structured JSON. Build your own UI or pipeline on top.

Licence

MIT