AI Audio Enhancement: Denoising, Remastering, Upscaling

Imagine a meeting recording at 8 kHz, mono, with constant hum and barely intelligible voices. Or an old archive cassette with crackle and hiss. This is standard for call centers and archives. With over 5 years in audio AI and 50+ completed projects, we build AI pipelines for audio enhancement that transform such material into clean speech in just days. Our AI audio enhancement services include audio upscaling, old recording restoration, audio denoising, mp3 artifact removal, and bandwidth extension. One client brought an 8 kHz, mono recording; we deployed a pipeline based on AudioSR and Resemble Enhance in three days. Results: PESQ from 1.8 to 3.9, STOI from 0.65 to 0.92. Voices became clear, high frequencies restored. Our client reduced subtitling costs by 50%. Contact us for a pilot project (starting at $2,000) – we'll process one of your recordings and show the result. Typical project cost ranges from $5,000 to $15,000.

Why Standard Methods Fall Short

Traditional equalizers and noise suppressors (FFmpeg anlmdn) work with the existing spectrum. They cannot recover frequencies lost during compression. The G.711 telephone codec cuts everything above 3.4 kHz – no filter can restore that data. AI models, in contrast, learn to reconstruct the spectrum from broadband speech datasets. AudioSR uses a diffusion probabilistic model to generate high-frequency components from scratch, performing spectral reconstruction and audio super-resolution. For bandwidth extension, this is the only working approach. Comparison: AI methods outperform traditional filters by 2–3 times in terms of PESQ score – typical AI improvement is 1.5–2.0 compared to 0.2–0.5 for DSP. Similarly, non-stationary noise removal is significantly better with neural networks.

How We Enhance Audio: From Analysis to Deployment

Our typical pipeline:

1. Analysis of source audio: spectrogram, noise level, codec, bitrate.
2. Model selection: AudioSR for upscaling, Resemble Enhance for denoising and mp3 artifact removal. AudioSR leverages a diffusion probabilistic model conditioned on low-frequency spectra; it learns the mapping to full-band audio via latent space.
3. Fine-tuning (optional): for specific domains (e.g., courtroom recordings), we do few-shot fine-tuning on 10–20 minutes of labeled data.
4. Integration: packaging into ONNX or Triton service, adding to processing stream.
5. Testing: metrics PESQ, STOI, SI-SNR, A/B test with three listeners.

Case Study: Restoring an Old Archive Recording (From Our Practice)

Task: a digitized cassette lecture – 16 kHz, 8-bit, mono, strong hiss and crackle. Our client wanted clean speech for subtitles.

We applied:

• AudioSR for upscaling to 48 kHz (restored frequencies up to 24 kHz).
• Resemble Enhance in denoise+enhance mode (removed hiss, improved clarity).
• FFmpeg for final loudness normalization (LUFS -16).

Results:

The entire process took 5 days. Our client received the pipeline code and documentation for independent deployment. Hear the difference in a demo – contact us, and we'll send a sample.

What's Included in Our Work

• Audio Analysis & Report: spectrograms, noise profiles, improvement potential.
• Model Selection & Fine-tuning: custom models for your domain.
• Pipeline Development: ready-to-deploy ONNX/Triton service with API.
• Documentation: installation guide, API reference, integration examples.
• Training: 2-hour session for your team on using and maintaining the pipeline.
• Support: 1 month of technical support after deployment.

Quality Metrics We Use

Primary objective metrics:

• PESQ (ITU-T P.862) – speech quality, target >3.5. Standardized by ITU-T.
• STOI – intelligibility, target >0.85.
• MOS-LQO – subjective quality, target >4.0.
• SI-SNR – signal-to-noise ratio, target >15 dB.

We guarantee a PESQ improvement of at least 1.0 point on your recordings. Measurements are taken before and after on a control set. Storage savings up to 30% after cleanup.

Guaranteed Results

For typical projects we deliver:

• PESQ increase from 1.8–2.0 to 3.5–4.0.
• STOI improvement from 0.65–0.75 to 0.85–0.95.
• ASR (speech recognition) error reduction by 30–50% after cleanup.
• Audio compression without quality loss – space savings up to 40%.

Use Cases

• Enhancing call center recordings before STT – recognition error drops by 30–50%.
• Preparing audio datasets for TTS fine-tuning – clean material without artifacts.
• Remastering archival materials (lectures, interviews) for streaming platforms.
• Preparing clean audio for training ASR models.

Neural networks for audio (AudioSR, Resemble Enhance) solve tasks beyond classical DSP. We bring 5+ years of audio AI experience and 50+ completed projects. Our stack: PyTorch, Hugging Face, ONNX Runtime. Get a demo pipeline for your recording – contact us. Attach audio samples, and we'll evaluate the project in 2 days and deliver a turnkey proposal. Typical project cost ranges from $5,000 to $15,000.

Speech Recognition and Synthesis: ASR, TTS, Voice Cloning

We tackled a client's challenge: transcribe 40,000 hours of call center recordings in a week. Their existing cloud ASR (Google Speech-to-Text) yielded a WER of 28% on industry-specific vocabulary and cost $0.006 per minute — prohibitively expensive at that volume. The goal was to reduce WER below 10% and switch to self-hosted inference. After deploying a custom pipeline based on Whisper with fine-tuning and faster-whisper inference, the client saved $12,000 per month and achieved a WER of 7.3%.

How does speech recognition ASR handle noisy call center recordings?

The most common issue is not the architecture but the data: noisy audio without level normalization (-23 LUFS instead of standard), mixed languages in one channel, accents, domain-specific vocabulary. Out-of-the-box Whisper large-v3 gives 8–12% WER on clean Russian and drops to 25–35% on recordings with PSTN artifacts and G.711 narrowband codec. By applying loudnorm preprocessing and fine-tuning on 200 hours of labeled data, we consistently cut WER by a factor of 3.

Typical problems we encounter

WER does not converge to the desired metric. Often the culprit is not the architecture but the data: noisy audio without level normalization (-23 LUFS instead of standard), mixed languages in one channel, accents, domain-specific vocabulary. Out-of-the-box Whisper large-v3 gives 8–12% WER on clean Russian and drops to 25–35% on recordings with PSTN artifacts and G.711 narrowband codec.

Diarization fails with more than two speakers. pyannote/speaker-diarization-3.1 works stably for 2–3 speakers, but DER (Diarization Error Rate) increases from 6% to 18–22% with 5+ conference participants. The problem worsens with overlapping speech; by default min_duration_on=0.1 cuts short interjections. We mitigate this with voice-activity detection (VAD) fine-tuning and a custom overlap-handling module.

Voice cloning — latency vs. quality. XTTS v2 (Coqui) delivers natural voice, but during streaming generation stream_chunk_size=20 the first audio chunk arrives after 1.4–2.0 seconds — unacceptable for interactive scenarios. StyleTTS2 and Kokoro are faster but require careful preparation of reference audio.

How do we solve it in practice?

The basic stack for a production pipeline:

• ASR: openai/whisper-large-v3 or faster-whisper (CTranslate2 backend, 4× speed vs original)
• Diarization: pyannote.audio 3.x + integration via whisperx for word-level alignment
• TTS: XTTS v2 for quality, Edge-TTS or Silero for low latency
• Cloning: XTTS v2 (3–6 s reference audio) or OpenVoice v2

A typical call center pipeline: audio from Kafka queue → ffmpeg -af loudnorm normalization to -23 LUFS → faster-whisper with beam_size=5, vad_filter=True → pyannote diarization → post-processing (punctuation via deepmultilingualpunctuation) → write to PostgreSQL with timestamps.

Case study from our practice. A fintech company with 12,000 calls per day. Initial WER on Russian with banking vocabulary — 22% (Google STT). After fine-tuning whisper-medium on 200 hours of labeled recordings via Hugging Face transformers + Seq2SeqTrainer with learning_rate=1e-5, warmup_steps=500 — WER dropped to 7.3%. Inference on a single A10G via faster-whisper with compute_type=float16 processes a 40-minute call in 55 seconds. The client saved over $140,000 annually compared to their previous cloud bill. Contact us for a free pilot estimate to see similar savings on your data.

How to fine-tune Whisper on domain data?

When a general model underperforms, fine-tuning is the first tool. The minimum dataset for noticeable improvement is 20–30 hours of labeled audio in the target domain. Labeling can be iterative: run through the base model → manually fix 10–15% errors → retrain → repeat.

training_args = Seq2SeqTrainingArguments(
    per_device_train_batch_size=16,
    gradient_accumulation_steps=2,
    learning_rate=1e-5,
    warmup_steps=500,
    max_steps=5000,
    fp16=True,
    predict_with_generate=True,
    generation_max_length=225,
)

Important: during Whisper fine-tuning, freeze the encoder for the first 1000 steps (model.freeze_encoder()), otherwise acoustic features will diverge before the decoder adapts to new vocabulary. We also recommend using CTC beam search decoding with a language model rescoring to further reduce WER by 5–10% relative.

Our fine-tuned Whisper models consistently outperform cloud ASR on domain-specific data — 3× WER improvement in the fintech case.

Speech synthesis: How to choose a model for your task?

For interactive bots requiring TTFB < 300 ms — Silero or Kokoro. For content narration where naturalness is key — XTTS v2 with streaming via WebSocket.

Our process and deliverables

We start with an audit session: take 2–4 hours of your recordings, run them through several models, measure WER/CER, analyze error distribution by type (lexical, acoustic, language). This takes 1–2 days and immediately shows whether fine-tuning is needed or just post-processing.

Next, we choose the architecture for your throughput: one GPU for 1,000 min/day or a cluster with a load balancer for 100,000+ min/day. Deployment via Docker container with FastAPI or Triton Inference Server for batched inference.

What you get after engagement:

• Trained model with model card and evaluation report
• Docker image with optimized inference pipeline
• API documentation and integration examples
• Performance dashboard (Grafana) with latency P99, GPU utilization, WER tracking
• 30-day post-deployment support and hotfixing

Timelines depend on complexity:

• Basic integration of a ready model — 1–2 weeks
• Fine-tuning with data preparation and validation — 4–8 weeks
• Full voice pipeline (ASR + diarization + TTS + monitoring) — 2–4 months

Project investments typically range from $20,000 to $80,000. Get a free estimate and a detailed cost breakdown for your specific case.

Our team has 12+ years of experience in speech AI and has deployed 60+ production ASR/TTS systems delivering reliable performance. Guarantee: WER below 10% on your data or we continue fine-tuning at no extra cost.

Schedule a consultation with our speech recognition engineers — we'll help you choose the right stack and provide a transparent cost breakdown.