When your downstream CRM integration logs the wrong customer name or misses a critical product entity, the failure rarely lies in your LLM prompt. It lies in the silent acoustic degradation of your 8kHz transcription layer. Modern speech AI models are trained and tested on clean, high-fidelity wideband audio (podcasts, audiobooks, conference recordings captured at 16kHz or higher), while legacy telephony infrastructure predominantly routes voice traffic through 8kHz narrowband codecs, creating a fundamental mismatch that no preprocessing alone can fully resolve.
The most common engineering response is upsampling: convert the 8kHz stream to 16kHz before passing it to a wideband model. This article explains precisely why that fails at the acoustic level, quantifies the accuracy loss you should expect in production, and gives you a concrete methodology for benchmarking your own telephony audio before committing to an STT infrastructure decision.
How narrowband codecs distort speech data
Narrowband constraints for STT accuracy
The public switched telephone network (PSTN) was designed around a single constraint: intelligibility for human listeners, not acoustic fidelity for machine learning models. The result is a narrowband bottleneck that limits transmitted audio to the 300 to 3400 Hz frequency range, a band wide enough for a human to follow a conversation but severely truncated for a model trying to distinguish phonemes.
The Nyquist-Shannon sampling theorem defines the upper bound precisely: an 8kHz sample rate can represent frequencies only up to 4kHz. Any spectral energy above that limit is entirely absent from the recorded signal, not degraded or distorted but physically uncaptured. Fricatives such as /s/ concentrate the bulk of their acoustic energy above 3,500 Hz, reaching as high as 10,000 Hz, well above the 4kHz Nyquist limit of an 8kHz signal. At 8kHz, none of that energy is captured.
Standard upsampling (linear interpolation) from 8kHz to 16kHz increases the sample rate in the file metadata but cannot regenerate spectral energy above the original Nyquist limit. The model receives a 16kHz file, but the meaningful spectral content above 4kHz remains absent from the upsampled input. Generic wideband models trained on real 16kHz audio will produce confused or incorrect outputs on fricative-heavy words because the acoustic features they were trained to detect are physically absent from the upsampled input.
How codec artifacts lower STT accuracy
The narrowband frequency limit is only the first layer of degradation. Telephony codecs apply compression algorithms on top of the already-limited signal, introducing quantization noise, phase distortion, and spectral smearing that compound the accuracy problem. Compression works by removing spectral redundancy that psychoacoustic models predict human listeners will not consciously perceive. The problem is that acoustic models for STT rely on precisely those fine-grained spectral details for phoneme discrimination: what a codec discards as perceptually irrelevant is often acoustically critical for automated transcription.
How packet loss degrades STT results
VoIP networks add a third degradation vector. Jitter and dropped packets are routine in production environments, and the standard mitigation is Packet Loss Concealment (PLC): an algorithm that generates a synthetic waveform to fill temporal gaps in the received stream, minimizing audible artifacts for human listeners.
Packet loss is a known, quantifiable driver of ASR error, and conventional concealment methods leave measurable word error rate on the table, which is why newer approaches replace the concealment step entirely rather than trying to patch it. The synthetic waveforms PLC inserts are optimized for perceptual smoothness, not acoustic fidelity to the original phonemes. A transformer-based acoustic model processes the PLC-generated segment as if it were a genuine phoneme sequence, producing confident but incorrect transcriptions at the output layer.
Quantifying accuracy loss in narrowband streams
Training set deficits in STT pipelines
The degradation problem is compounded by how most commercial and open-source STT models are trained. Generic wideband models learn phoneme discrimination from large-scale datasets dominated by clean, high-sample-rate recordings: read speech, broadcast audio, podcasts, and audiobooks. Telephony audio, with its narrowband filtering, codec compression, and network artifacts, represents a fundamentally different acoustic distribution. A model trained almost entirely on clean 16kHz data develops representations that do not generalize to the acoustic features present in degraded 8kHz signals. A distribution mismatch confirmed by research on mixed-bandwidth training, which found that wideband ASR models do not perform well on narrowband speech data.
How narrowband filtering degrades transcription accuracy
On real-world telephony recordings with additional codec compression and network artifacts, accuracy drops substantially further below clean-audio baselines.
How 8kHz transcription errors compound downstream
Raw transcription error rate understates the business impact because errors do not fail uniformly across word categories. While function words (articles, prepositions, short connectors) account for the most frequent individual errors, the business impact concentrates on content words: the named entities, product terms, and numerical identifiers your downstream NLU pipeline depends on. A key data extraction pipeline that corrupts customer account numbers or misidentifies product references will degrade CRM data quality and coaching scorecard reliability far beyond what the aggregate error rate suggests. The failure is silent at the transcript level but immediately visible in automated call disposition accuracy and QA routing workflows.
Codec limitations in 8kHz speech to text pipelines
Narrowband signal loss in G.711 streams
G.711 is the foundational PSTN codec. It uses logarithmic companding (mu-law in North America and Japan, a-law in Europe) to encode speech at 64 kbps, applying less aggressive compression than later codecs while still hard-limiting the representable frequency range to 4kHz. G.711 is the best-case scenario for telephony STT: minimal compression distortion, preserved amplitude fidelity within the captured band, but every fricative peak above 4kHz is permanently absent regardless of source audio quality.
Codec effects: G.729 vs. Opus
G.729 compounds the narrowband frequency limitation with aggressive Linear Predictive Coding (LPC), encoding 8kHz speech at 8 kbps. The algorithm models the vocal tract filter to predict the speech signal, then transmits only the residual error. What LPC discards as acoustically predictable often includes the fine spectral texture that STT acoustic models use for speaker-specific phoneme discrimination and diarization. The perceptual quality reduction is measurable, and the impact on diarization is more severe than on transcription alone, because speaker identity information sits in the spectral characteristics the codec de-emphasizes.
Opus offers variable bitrate with wideband capability but is commonly constrained to narrowband mode in legacy telecom deployments. Where Opus operates in wideband mode at adequate bitrates, it substantially outperforms G.729 for STT purposes. The practical challenge is that you often cannot control or verify the codec negotiated end-to-end in a production SIP (Session Initiation Protocol) trunk.
Why AMR-NB degrades STT accuracy
Adaptive Multi-Rate Narrowband (AMR-NB) is the dominant codec in GSM mobile networks, operating across eight codec modes from 4.75 kbps to 12.2 kbps. The codec adjusts compression aggressiveness in real time based on network channel quality, creating recordings where acoustic quality fluctuates on a frame-by-frame basis (every 20ms). The resulting quality variance within a single recording is harder for acoustic models to handle than a consistently lower-quality signal, because the model's learned representations cannot adapt to shifting spectral characteristics mid-call.
Table 1: Codec comparison and STT impact
| Codec |
Bitrate |
Primary distortion mechanism |
Impact on STT and diarization |
| G.711 |
64 kbps |
Hard 4kHz frequency cutoff, logarithmic companding |
Fricatives permanently lost above 4kHz, otherwise best-case narrowband fidelity |
| G.729 |
8 kbps |
LPC discards non-speech spectral features |
Significant degradation to speaker identity features needed for diarization |
| AMR-NB |
4.75-12.2 kbps |
Dynamic bitrate switching on network conditions |
Quality variance within a single recording due to dynamic codec switching. Acoustic models face inconsistent spectral conditions that do not match a stable learned representation |
| Opus (narrowband) |
Variable |
Mode-constrained to 8kHz in legacy deployments |
Similar to G.711 when bandwidth-constrained, significantly better in wideband mode |
Quantifying packet loss impact on STT accuracy
Production loss rates for STT pipelines
Well-managed enterprise VoIP networks target below 1% packet loss, but production conditions frequently exceed this in practice. Acceptable packet loss thresholds for voice communications are generally defined at below 1% for high-quality calls, with anything above 5% considered high loss. Mobile networks during handoff events or under load can produce transient bursts well above these thresholds. For a two-minute phone call sampled at 8kHz, even a brief loss event means some fraction of audio is reconstructed by a PLC algorithm rather than captured from the actual speaker.
Why packet loss degradation is non-linear
The degradation pattern under packet loss is not linear. As loss rates climb, the proportion of PLC-generated audio increases, and the density of synthetic artifacts in the stream grows non-linearly because PLC artifacts degrade the acoustic context that the model uses to anchor its predictions for adjacent frames. Telephony-optimized models trained on degraded audio distributions handle these conditions more reliably because they have learned to recognize phonemes despite PLC-characteristic spectral patterns.
Why signal recovery fails on 8kHz
Standard jitter buffer algorithms and packet recovery techniques are designed to maintain perceptual continuity for human listeners, not to preserve the fine acoustic details that phoneme recognition requires. When a PLC algorithm fills a 20ms gap in a narrowband 8kHz stream, it generates a synthetic waveform optimized to minimize perceptual artifacts as measured by Mean Opinion Score (MOS), a human listener quality rating, not a measure of acoustic fidelity to the original phoneme sequence.
The acoustic model, however, is trained on real phoneme transitions: the spectral trajectory of the PLC-generated segment will not match any genuine phoneme sequence in the model's learned representations, producing confident but incorrect transcriptions. This is precisely why researchers have proposed front-end adaptation networks that replace traditional PLC entirely, substituting the concealment step with a learned network optimised for ASR robustness rather than perceptual smoothness. Current production PLC algorithms were not designed with ASR as the downstream consumer.
Benchmarking STT for 8kHz narrowband audio
1. Collect representative telephony samples
The most common benchmarking mistake is testing on clean audio. Your evaluation set must include recordings from your actual production pipeline, covering the full distribution of accents, background environments, overlapping speech, and call quality you will encounter in production. A model that performs well on sanitized test sets can show a 10 to 20 percentage point accuracy gap on real call center audio. For multilingual contact centers handling global CX, your evaluation set must also cover each language and accent distribution separately, because performance gaps are not uniform across languages.
2. Assess codec-induced signal loss
Identify the codecs active in your production telephony stack before testing. Most SIP trunking providers log the codec negotiated per call, and many call recording platforms expose this metadata. If you are working with WebRTC-based contact center infrastructure such as Amazon Connect or Vonage, inspect the session description protocol (SDP) negotiation logs. Grouping evaluation samples by codec allows you to quantify which codec is driving the largest accuracy gap.
3. Model network-induced packet loss
Use NetEm (Linux kernel network emulator) to add controlled packet loss and jitter to your test audio pipeline. A structured test should evaluate your chosen model at 0%, 1%, 5%, and 10% packet loss rates to characterize the degradation curve. The gap between 0% and 5% loss shows the accuracy impact of modest packet loss typical in many production environments.
4. Measure accuracy in degraded environments
Calculate word error rate on degraded samples using reference transcripts that include named entities, numerical identifiers, and domain-specific vocabulary at representative density. Aggregate error rate will understate the business impact: calculate entity-level error rates separately to quantify how often the model corrupts the tokens your downstream NLU pipeline depends on. For multi-speaker recordings, also measure diarization error rate using speaker diarization on your async pipeline.
5. Evaluate your live telephony mix
Once you have a quantified baseline on your internal evaluation set, run a blind cross-provider evaluation. The blind STT comparison tool strips out provider branding so you pick the better transcript before seeing which vendor produced it, removing the confirmation bias that corrupts most internal vendor evaluations. It is a useful gut check, but follow it with a reproducible benchmark on your full audio distribution before making a production commitment.
Key metrics for assessing telephony STT models
Optimizing STT for 8kHz audio sources
The only technically sound approach to telephony STT is a model natively trained on 8kHz audio data, not a wideband model applied to upsampled input. When we train models natively on narrowband data, they develop phoneme representations that function within the 300 to 3400 Hz frequency band, without relying on spectral features above 4kHz that telephony recordings never capture. Models fine-tuned on downsampled versions of wideband training data still underperform models trained on genuine telephony recordings, because downsampled data does not contain the codec compression artifacts, PLC-generated segments, or dynamic quality variation present in real call center audio.
Optimizing 8kHz audio for accuracy
Pre-processing steps like noise suppression and automatic gain control can reduce background noise and level variation present in telephony recordings, and they are worth including in your pipeline. The important constraint is that they operate on audio that arrives at 8kHz: they can reduce noise within the representable frequency band, but they cannot recover spectral energy above the Nyquist limit that the codec never captured. Treat pre-processing as a useful complement to a telephony-optimized model, not as a substitute for one.
Quantifying narrowband transcription accuracy
The downstream NLU performance of your pipeline is the primary KPI, and raw transcription error rate is only a proxy for it. The metrics that map most directly to business outcomes are F1-score on entity extraction (does the model correctly identify customer names, account numbers, and product terms), intent classification accuracy on your actual call taxonomy, and call disposition automation accuracy across your QA and routing workflows. A model that reduces aggregate error rate by three percentage points but concentrates its remaining errors on named entities will degrade your NLU pipeline more than a model with slightly higher aggregate error rate but more uniform error distribution.
Table 2: Build vs. buy decision matrix for telephony STT
| Evaluation criteria |
Self-hosted open-source |
Specialized telephony API (Solaria-3) |
| Initial time to staging |
Significant engineering effort (GPU provisioning, model selection, pipeline integration) |
Sub-24 hours via REST API |
| Ongoing maintenance overhead |
GPU ops, model versioning, scaling, accuracy monitoring |
Vendor-managed, monitored via status.gladia.io |
| Accuracy on noisy/accented 8kHz |
Variable, often significantly higher error rates in production than staging benchmarks |
#1 on Switchboard ahead of AssemblyAI, ElevenLabs, Deepgram, Mistral, and Speechmatics |
| Diarization quality |
Requires separate integration and additional model maintenance |
High-quality speaker diarization, included at base rate |
| Predictable unit economics |
Variable GPU cost scales with concurrency |
As low as $0.20/hr async on Growth, all features included |
Addressing 8kHz telephony transcription accuracy
Expected WER for 8kHz telephony audio
Our Solaria-3 model, built specifically for real-world business and conversational audio, ranks #1 on Switchboard (the most challenging conversational telephony dataset) ahead of AssemblyAI, ElevenLabs, Deepgram, Mistral, and Speechmatics, with a 33.9% WER (the only model under 35%). It also leads the same field on the Earnings22 financial calls benchmark.
Gravite, a French CCaaS quality-monitoring platform, processes 50,000 hours of audio per year through our pipeline and cut call review time 93%, reducing per-call review from roughly 15 minutes to 1 minute with no fallback provider required.
Codec efficiency in lossy networks
We trained Solaria-3 on real-world business audio that includes the acoustic artifacts introduced by codec compression and network degradation. The model learns to recognize phonemes in their degraded forms, including the spectral smearing characteristic of G.729, the variable quality profiles introduced by AMR-NB switching, and the temporal gaps created by PLC algorithms. This is not post-processing or compensation: the model's acoustic representations are built from the same distribution of degraded audio your production pipeline will deliver.
Testing STT quality on static audio
For teams that want to validate before writing any integration code, the Gladia CLI runs a complete transcription with diarization on a folder of telephony recordings in a single terminal command. Install via curl -fsSL https://github.com/gladiaio/gladia-cli/releases/latest/download/install.sh | sh on macOS or Linux.
How Gladia handles 8kHz telephony
Our end-to-end pipeline covers the full async stack: transcription with Solaria-3, speaker attribution through speaker diarization, named entity recognition, text-based sentiment analysis, and summarization, all delivered through a single API call with no separate enrichment vendor to maintain. The audio intelligence feature suite is included at the base rate on Starter and Growth plans. Speaker attribution can be handled in post-processing for higher accuracy on completed call recordings, which is the standard architecture for CCaaS quality monitoring and post-call intelligence platforms.
For teams integrating for the first time, npx skills add gladiaio/skills installs Gladia Skills into your AI coding agent (Cursor, Claude Code), giving it accurate context on the API and SDK surface to avoid hallucinated parameters during implementation.
Compliance and data residency checklist for telephony deployments:
- SOC 2 Type II: Audited and current, covering dedicated cloud clusters in EU and US regions
- GDPR: EU data residency by architecture, documented in our DPA
- HIPAA: Certified, covering applicable data handling requirements for healthcare-adjacent deployments
- ISO 27001: Information security management standard, audited
- Data training policy: On Growth and Enterprise plans, customer audio is never used for model training, no opt-out required. On the Starter plan, data can be used for training by default
- PII redaction: Available as an audio intelligence feature but must be explicitly configured, not enabled by default
Growth plan pricing starts as low as $0.20/hr for async transcription with diarization, sentiment analysis, named entity recognition, and summarization all included at the base rate.
Start with €50 in free credits on the Starter Plan to evaluate against your own telephony audio.
FAQs
What is the expected accuracy of Solaria-3 on telephony audio?
Solaria-3 ranks #1 on Switchboard, the most challenging conversational telephony benchmark, ahead of AssemblyAI, ElevenLabs, Deepgram, Mistral, and Speechmatics, with a 33.9% WER (the only model under 35%).
Does Gladia support real-time speaker diarization?
No, speaker diarization is powered by pyannoteAI and is strictly available in (batch) async workflows only. For real-time transcription pipelines, speaker attribution can be handled in post-processing for higher accuracy on completed recordings.
How much do Gladia's audio intelligence features cost?
Diarization, sentiment analysis, and named entity extraction are included in the base rate: $0.61/hr async on the Starter plan and as low as $0.20/hr async on the Growth plan, with no per-feature add-on charges.
Why does upsampling 8kHz audio to 16kHz fail to fix transcription accuracy?
Standard upsampling (linear interpolation) changes the sample rate in the file metadata but cannot regenerate spectral energy above the original Nyquist limit. Because the source recording at 8kHz captures frequencies only up to 4kHz, the fricative energy above 3,500 Hz that generic wideband models rely on for phoneme discrimination is permanently absent from the signal. The upsampled file fills the frequency space above 4kHz with values interpolated from the existing samples, but because the original 8kHz recording captured no spectral energy above the Nyquist limit of 4kHz, those interpolated values do not represent genuine phoneme content from the original speech signal. The wideband model receives a 16kHz file with no new phonemic information above 4kHz.
What are realistic packet loss rates in production VoIP pipelines?
Well-managed enterprise VoIP networks target below 1% packet loss, with anything above 5% classified as high loss. Mobile networks under load or during handoff events can produce transient bursts that significantly exceed these thresholds, introducing PLC-generated audio segments that STT acoustic models were not trained to handle.
Key terms glossary
Narrowband: Audio limited to the 300 to 3400 Hz frequency range, characteristic of traditional telephony networks and an 8kHz sample rate.
Nyquist frequency: Half of a system's sampling rate, defining the maximum frequency that can be accurately represented. For an 8kHz signal, the Nyquist limit is 4kHz.
Fricative: A consonant sound produced by forcing air through a narrow channel, such as /f/, /s/, and /th/, whose acoustic energy, in the case of /s/, concentrates above 3,500 Hz and can reach as high as 10,000 Hz, well above the 4kHz Nyquist limit of an 8kHz signal.
Diarization: The process of partitioning an audio stream into segments according to speaker identity, determining who spoke at each point in the recording.
Packet Loss Concealment (PLC): A masking technique used in VoIP networks to generate synthetic audio segments that cover dropped voice packets, maintaining perceptual continuity for human listeners.
Linear Predictive Coding (LPC): A compression method used in codecs like G.729 that models the vocal tract filter and transmits only the residual error, discarding spectral features the algorithm predicts are acoustically redundant.
Mean Opinion Score (MOS): A subjective perceptual quality rating for audio codecs, scored from 1 (bad) to 5 (excellent), reflecting human listener perception rather than machine transcription accuracy.
